{"id":"46f3df24-9911-4bfd-83cc-33244248b031","arxiv_id":"2411.13716","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper constructs normative joint-angle ranges across the gait cycle from 2D pose estimates and uses one-standard-deviation bands to flag abnormal gait.","lead":"This paper builds average joint-angle curves across 351 normal gait cycles from three RGB video datasets and uses them as a reference for flagging abnormal moments in new walking videos. It aims to give clinicians a low-cost, video-only way to see which joints deviate from typical gait and when.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The clinical-norm claim depends on pooling uncalibrated 2D joint angles across datasets with different camera geometries; no evidence yet shows these angles are comparable enough to form one normative population.","rationale":"The paper's central claim is that the derived parameters are clinically representative normative gait values. For that claim to hold, the 2D joint-angle measurements pooled across datasets must measure the same quantity in comparable units. Equation (1) computes angles from pixel coordinates in uncalibrated monocular video; without calibration, the same 3D joint angle projects to different 2D angles depending on camera placement, height, and lens distortion. Pooling across three datasets with different capture conditions therefore risks creating a mixture distribution, and the one-standard-deviation abnormality threshold has no clear clinical interpretation if the pooled mean is not a single population norm. The test videos in Section V come from the same datasets used to build the norms, so the experiments demonstrate internal consistency at best, not external clinical validity. The reader's conditional verdict is appropriate: the method is clearly described, cycle normalization is standard practice, and the authors acknowledge some limitations, but the normative claim requires either per-dataset comparability evidence or external validation against established clinical or marker-based measurements. I do not see grounds to reject outright, because the pipeline is reproducible in principle and the limitation section is candid. The concrete test of per-dataset mean curves directly settles whether pooling is justified; the sample-count discrepancy in Section IV.A versus Table I should first be resolved so the test runs on the correct data.","tokens_in":9877,"tokens_out":4638,"duration_ms":48860,"concrete_test":"For each joint, compute the mean gait-cycle angle curve separately for GPJATK, GAVD, and CASD using the same interpolation and normalization. Run a permutation test of the null hypothesis that the three dataset-specific mean curves are equal, using the maximum absolute between-curve difference as the test statistic and permuting dataset labels among cycles to build the null distribution. If the observed differences exceed the within-dataset spread (95% CI excludes zero), the pooled 351-cycle norm is an artifact of camera/dataset differences and the clinical-norm claim fails. Also reconcile the Table I versus Section IV.A sample-count discrepancy by re-inspecting the original datasets before interpreting any result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The normative curves are built by pooling joint angles computed from 2D BlazePose keypoints via Eq. (1) across GPJATK, GAVD, and CASD (Section IV.E, Table I). These are not anatomical angles: with a single lateral view and no camera calibration (acknowledged in Section V-C), the measured 2D angle depends on camera height, pan, distance to the subject, and lens distortion, not only on joint kinematics. If camera geometry differs across the three datasets, the pooled mean and SD curves are a mixture of dataset-specific offsets rather than a clinically meaningful norm. Section V's abnormality flags (points outside one SD) would then partially flag dataset artifacts instead of gait deviations. The paper provides no per-dataset mean curves, no between-dataset variance comparison, and no validation against marker-based motion capture. The internal inconsistency in sample counts (Section IV.A text says GPJATK: 42 subjects/99 cycles and GAVD: 11 subjects/71 cycles, while Table I swaps these values and reports CASD cycles as 181 vs 137 in the text) further weakens confidence that the pooled dataset is accurately described. Unless the 2D angles are shown to be comparable across datasets, the central claim of clinically representative normative values is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven pipeline for building normative gait-cycle joint-angle parameters from monocular RGB video, using BlazePose keypoints and a 2D vector-based angle computation (Eqs. 1–2). The authors pool 351 'normal' gait cycles from three datasets (GPJATK, GAVD, CASD), temporally normalize them to percent gait cycle, interpolate with cubic splines, and compute per-joint mean and standard-deviation bands as normative values. They then flag any angle outside one standard deviation as potentially abnormal, visualize single- and multi-joint deviations, and demonstrate the approach on four videos, claiming cycle-wise kinematic deviation and abnormality detection for clinical use.","tokens_in":10105,"tokens_out":2722,"duration_ms":26021,"significance":"If the normative parameters were properly validated, the approach could provide a low-cost, accessible alternative to marker-based gait analysis, with the practical advantage of using only a single RGB camera. The use of joint angles—an established clinical kinematic measure—and the emphasis on cycle-wise, phase-localized abnormality detection are commendable. The paper also ships a clear pipeline description that is easy to reproduce from the text. However, the clinical significance claimed by the authors is presently unsupported: the normative values are unvalidated sample statistics from a small, heterogeneous, and uncalibrated dataset, and the evaluation is qualitative and circular. The contribution is best read as a baseline framework, not as clinically usable normative values.","major_comments":[{"comment":"The central claim of 'clinically representative normative values' rests on pooling uncalibrated 2D joint angles from videos captured with different camera geometries across three datasets. With a single lateral view and no camera calibration (acknowledged in Section V-C), the 2D angle from Eq. (1) depends on camera height, pan, and distance, not only on joint kinematics. If camera geometry differs across datasets, the pooled mean and standard deviation curves are a mixture of dataset-specific offsets. The paper provides no per-dataset mean curves, no between-dataset variance comparison, and no validation against marker-based motion capture. Without evidence that these 2D angles are comparable across datasets, the normative values do not support the clinical claim.","section":"Section IV.E / Table I / Section V-C"},{"comment":"The four test videos are drawn from the same three datasets that define the normative bands (Table I vs. Table III), which makes the evaluation circular: a 'typical' test video largely falls near a mean it helped create. More seriously, test video four uses subject 3 from CASD, and the same subject's normal gait cycles contribute to the normative pool, so the same individual is used to define normality and then evaluated as atypical. The reported detection results therefore conflate within-subject and between-subject variability and cannot demonstrate abnormality detection against a normative population. The authors should evaluate on held-out subjects or an external dataset.","section":"Section V.A / Table III / Section IV.E"},{"comment":"The abnormality detection threshold is set to one standard deviation from the mean with no clinical justification or sensitivity analysis. Under Gaussian assumptions, this threshold flags roughly 32% of a normal population's cycles as 'abnormal' by construction, so the observed red dots are not evidence of gait abnormality. The paper does not show that one standard deviation corresponds to a clinically meaningful threshold, nor does it test how the detection output changes with alternative thresholds. Without such analysis, the claimed ability to 'detect potential deviations' is not established.","section":"Section V.B / Fig. 2"},{"comment":"The experimental evaluation is entirely qualitative: four videos, visual inspection of red/blue dots, and no quantitative accuracy metrics, no statistical comparison between typical and atypical conditions, and no comparison to a reference standard. The claim that 'atypical gait patterns' produce 'an observable increase' in deviations (Section V.B.2) is supported only by eyeballing Figures 2–6. The paper needs quantitative evaluation, e.g., sensitivity/specificity for classifying held-out typical vs. atypical cycles, or a per-cycle deviation score with a statistical test.","section":"Section V.B / Section V-C"},{"comment":"The sample counts are internally inconsistent. The text states GPJATK has 42 subjects/99 cycles and GAVD has 11 subjects/71 cycles, while Table I swaps these values (GPJATK 11/71, GAVD 42/99). In addition, the text reports 137 complete normal gait cycles for CASD, but Table I lists 181. These discrepancies are load-bearing because the normative curves and any pooled statistics depend on the exact composition of the dataset; they must be resolved and the counts verified against the original dataset documentation.","section":"Section IV.A / Table I"}],"minor_comments":[{"comment":"The authors describe the normative pool as 'a large RGB video dataset of typical gait cycles,' but the dataset comprises 73 subjects and 351 cycles, many from the same subjects and captured in controlled conditions. 'Large' is an overstatement for a clinical-norm claim; consider a more modest description such as 'a multi-dataset pool of 351 normal cycles.'","section":"Section IV.E"},{"comment":"The method assigns each frame a position within the gait cycle, but the text does not describe how the cycle percentage is computed from the annotated heel-strike events. It would clarify the pipeline to state whether linear interpolation between annotated start and end frames or a more sophisticated phase estimation is used.","section":"Section IV.E"},{"comment":"The caption says 'outside of standard deviation of mean' but should read 'outside one standard deviation of the mean' for consistency with Fig. 11 and the main text.","section":"Fig. 12 caption"},{"comment":"The dataset name is inconsistently written as 'GAVD' in most places but 'GA VD' in Sections IV.A.2 and Fig. 8; unify the spelling.","section":"Throughout"},{"comment":"The limitation paragraph describes the dataset as 'uncontrolled in nature,' but GPJATK and CASD are controlled indoor recordings per Section IV.A. This contradicts the earlier description and should be reworded to refer only to the GAVD component or to the overall heterogeneity.","section":"Section V-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a promising idea—building normative gait parameters from RGB video—but the current evidence is insufficient for the clinical claims. The circular evaluation (same subjects in the norm and test sets) and the uncalibrated pooling problem are likely fixable with additional experiments and a held-out subject protocol, so I do not recommend rejection. However, the internal inconsistency in Table I must be corrected, and the threshold choice must be justified. I would also advise the editor to consider whether the paper's length and evaluation depth match the journal's standards for a clinical claims paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: it's a reproducible pipeline for building joint-angle waveforms from monocular video, but the clinical-norm claim outruns the evidence. The method itself is fine; the validation is not.\n\nWhat's actually new is modest: pooling three datasets (including the authors' own CASD) and applying standard pose-estimation plus vector-angle math to emit mean/SD curves. They also provide a useful multi-joint visualization for flagging deviations. The writing is straightforward, and the limitations section is honest about the lack of camera calibration.\n\nThe soft spots are real. The biggest is pooling uncalibrated 2D angles from datasets with different camera geometry. A 2D joint angle from a single lateral view depends on camera height, pan, and distance, not just kinematics. If the three datasets have different setups, the pooled norm is a mixture of offsets, and your deviation flags become dataset artifacts. The paper gives no per-dataset curves or between-dataset variance comparison, so this concern lands.\n\nSecond, the abnormality detection is circular: the four test videos come from the same three datasets that built the norms. Video four is the same subject who contributed normal gait to the norm and is then evaluated as atypical. That doesn't demonstrate detection of abnormality; it demonstrates distance from a mean the subject helped create.\n\nThird, there's an internal inconsistency in the sample counts: the text gives GPJATK 42 subjects/99 cycles and GAVD 11/71, but Table I swaps these and reports CASD as 181 cycles vs 137 in the text. Either way, the pooled norm is built from a small, heterogeneous set with multiple cycles per subject treated as independent, which further weakens the 'normative population' claim.\n\nFourth, the one-SD threshold is arbitrary and the evaluation is qualitative on four videos, with no comparison to marker-based motion capture or clinical ground truth.\n\nNone of this is fatal to the underlying idea. The paper would be a reasonable baseline if reframed as 'a demonstration of a video-based kinematic deviation flagging tool with known limitations.' As it stands, the abstract and conclusions overclaim clinical validity.\n\nI'd send it to review because it's reproducible and the topic matters, but I'd expect major revision: fix the sample counts, add per-dataset breakdowns, validate against a gold standard, and temper the normative claims. For a reading group, it's a decent case study on the pitfalls of pooling convenience data.","headline":"A clean pipeline for video-based joint-angle norms, but the clinical-norm claim outruns the evidence; worth review with major revision.","tokens_in":10653,"tokens_out":2115,"would_cite":false,"duration_ms":18901,"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":"This paper derives normative joint-angle parameters for the human gait cycle from 351 RGB-video cycles of typical gait, and uses them to flag kinematic abnormalities in single and multiple joints.","keywords":["gait analysis","human pose estimation","normative gait parameters","joint angles","clinical gait analysis","RGB video","kinematic deviation","gait cycle"],"falsifier":"Concrete test: record a separate group of healthy adults from the same left-side view, run the same pose-estimation and joint-angle pipeline, and count how many of their cycles fall outside the one-standard-deviation band. If a large fraction of known-normal cycles are flagged as abnormal, the normative band is too tight or the pooling assumption is wrong. A second decisive check is to compare the 2D normative curves against marker-based 3D motion capture on the same subjects; if the 2D-to-3D discrepancy is comparable to the width of the normal band, the band reflects camera geometry rather than biological variation.","tokens_in":9644,"feed_emoji":"🏃","tokens_out":5937,"duration_ms":45747,"temperature":0.7,"pith_summary":"This paper attempts to show that a usable clinical norm for human gait can be built purely from monocular RGB video, without markers or wearable sensors. The authors pool 351 typical gait cycles from three video datasets, extract 2D joint angles for ten joints with a lightweight pose estimator, and normalize every cycle from heel strike to heel strike. They then define the normative parameter for each joint as the mean and standard deviation of these angles across the pooled cycles. On four unseen videos, joint angles that fall outside one standard deviation from the mean are flagged as abnormal, for single joints and multiple joints simultaneously. The value of the claim is that a clinician could screen gait deviations with an ordinary camera and get a temporally localised, explainable readout.","feed_headline":"Gait norms from 351 RGB video cycles flag abnormal joints","feed_subtitle":"A mean and standard deviation per joint across the gait cycle lets clinicians screen movement with a single camera.","key_machinery":"The central object is the per-joint normative curve: for each of ten joints, a mean angle and a one-standard-deviation band across the 0–100% gait cycle, computed from 351 temporally normalised cycles. The mechanism that carries the argument is the vector-based joint angle calculation using the arctan2 function on the 2D keypoint coordinates, followed by cubic spline interpolation to align every cycle to a common time base. The normative mean and standard deviation serve as the reference distribution; a joint angle in a test video is deemed potentially abnormal when it lies outside one standard deviation of the mean at the corresponding cycle percentage. This statistical definition is what turns raw pose estimates into clinically interpretable flags.","core_discovery":"The paper's central claim is that a clinically representative normative parameter for each of ten joints can be derived from a large set of typical gait cycles recorded in RGB video, and that this parameter supports cycle-wise kinematic deviation and abnormality detection. The normative curves are built by computing 2D joint angles from BlazePose keypoints, mapping each frame to a percentage of the gait cycle bounded by the left heel strike, and interpolating missing frames with cubic splines. The publication states that these mean and standard deviation curves give clinicians reference values to compare a patient's joint angles against a normative population using only monocular video. Demonstrations on four test videos show that atypical gait produces more detected abnormalities than typical gait, and that the degree of deviation can be visualised as shading over the cycle.","pith_inferences":["If the pooling assumption holds, the same pipeline could build norms for other actions, since joint angles are action-agnostic and the cycle definition could be adapted.","A validation study against clinician ratings on a larger set of videos would likely tune the one-standard-deviation threshold, since 'abnormal' in clinical practice is not a fixed statistical distance.","Because multiple cycles come from the same subjects, the reported standard deviations probably understate between-subject variability; a random-effects model treating subject as a cluster would give a more honest norm.","The left-side-view-only construction means the norms are viewpoint-specific; a direct test would be to compute norms from one dataset and evaluate abnormality detection on the others without pooling."],"forward_implications":["A clinician could screen a patient's gait from a single lateral video and immediately see which joints deviate from the norm and at which phase of the gait cycle.","Because the analysis is cycle-wise, the same pipeline can track a patient over time, comparing each new cycle against the same normative band to monitor change during rehabilitation.","Joint angles are a universal kinematic measure, so the normative-curve approach is not locked to gait; the same machinery could be applied to other actions once cycles are defined.","The severity visualisation gives a prioritisation cue: darker regions indicate larger deviations, which can help clinicians focus attention on the most affected joints.","This establishes a baseline normative reference where none existed for RGB-video-only gait analysis, which is a prerequisite for automated abnormality screening."],"supporting_citations":[{"why":"Supplies the GPJATK videos used as one of the three normal-gait sources for building the normative curves.","marker":"[33]"},{"why":"Supplies the GAVD and CASD videos, including the clinically screened normal and atypical gait sequences used for both norm-building and testing.","marker":"[32]"},{"why":"The BlazePose pose estimator that extracts the 2D keypoints from which joint angles are computed.","marker":"[37]"},{"why":"The vector-based arctan2 method used to compute joint angles from three keypoints.","marker":"[40]"},{"why":"Cubic spline interpolation applied to handle missing frames and produce smooth joint-angle curves over the cycle.","marker":"[41]"},{"why":"The Conventional Gait Model, cited as the basis for using mean and standard deviation as normative values and tolerances in gait analysis.","marker":"[43]"},{"why":"Validation of BlazePose as a viable tool for gait analysis, which the paper relies on to justify the choice of pose estimator.","marker":"[38]"}],"fun_headline_variants":[],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that 351 gait cycles pooled from three video datasets by different cameras and subjects, with multiple cycles from the same subject, can be treated as one exchangeable population of normal gait, so the pooled mean and standard deviation curves represent a clinically meaningful norm.","fun_headline_variants_meta":{"error":"'choices'"},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:57:34.138644+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Concrete test: record a separate group of healthy adults from the same left-side view, run the same pose-estimation and joint-angle pipeline, and count how many of their cycles fall outside the one-standard-deviation band. If a large fraction of known-normal cycles are flagged as abnormal, the normative band is too tight or the pooling assumption is wrong. A second decisive check is to compare the 2D normative curves against marker-based 3D motion capture on the same subjects; if the 2D-to-3D discrepancy is comparable to the width of the normal band, the band reflects camera geometry rather than biological variation.","supporting_citations":[{"cited_title":"Calibrated and synchronized multi-view video and motion capture dataset for evaluation of gait recognition,","cited_arxiv_id":null,"evidence_quote":"Supplies the GPJATK videos used as one of the three normal-gait sources for building the normative curves."},{"cited_title":"Blazepose: On-device real- time body pose tracking,","cited_arxiv_id":null,"evidence_quote":"The BlazePose pose estimator that extracts the 2D keypoints from which joint angles are computed."},{"cited_title":"Sports2d - angles from video,","cited_arxiv_id":null,"evidence_quote":"The vector-based arctan2 method used to compute joint angles from three keypoints."},{"cited_title":"27 of Applied Mathematical Sciences, Springer-Verlag, New York, 01 1978","cited_arxiv_id":null,"evidence_quote":"Cubic spline interpolation applied to handle missing frames and produce smooth joint-angle curves over the cycle."},{"cited_title":"1–19, Springer International Publishing, 05 2017","cited_arxiv_id":null,"evidence_quote":"The Conventional Gait Model, cited as the basis for using mean and standard deviation as normative values and tolerances in gait analysis."},{"cited_title":"Validation of markerless 3d human pose estimation using blazepose for gait analysis,","cited_arxiv_id":null,"evidence_quote":"Validation of BlazePose as a viable tool for gait analysis, which the paper relies on to justify the choice of pose estimator."}],"review_version":1}