{"id":"55e3ced3-434e-4f82-9d72-d0a5871965fc","arxiv_id":"1908.07415","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A sparse deep autoencoder trained on normal gait skeletons produces a per-skeleton gait abnormality index, and sequence averaging of a weighted reconstruction error reaches 0.945 AUC on a nine-subject simulated dataset.","lead":"This paper uses three small autoencoders trained only on normal walking skeletons to score how abnormal a new skeleton is by its reconstruction error. A weighted sum of the three errors, averaged over a gait sequence, separates normal from simulated abnormal gaits on a nine-subject treadmill dataset.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-sequence results come from one split of only 4 test subjects, and Table 2 itself shows a tie (0.139) with the one-class SVM baseline, so the 'best ability' claim is not established.","rationale":"The reader correctly identifies the evaluation as a weak point, and I agree that no error bars and a single split make the central claim fragile. My concern sharpens this: because the 36 per-sequence scores come from only 4 subjects, the effective sample size is 4, and Table 2 already shows a tie with the one-class SVM at per-sequence error 0.139. This is an internal contradiction with the 'best ability' claim, not merely an external robustness worry. I did not make the per-frame normalization assumption the primary attack because the paper's own per-frame results (AUC 0.855) suggest some postural signal survives normalization; the more immediate and decisive problem is that the reported comparison cannot statistically distinguish the proposed weighted index from alternatives. The weight formula in Equation (2) is also ambiguously normalized, but since a constant scaling of weights does not change threshold-based classification, it is less load-bearing than the statistical fragility. A leave-one-subject-out evaluation would settle whether the weighted-sum advantage is real; if it disappears, the central claim should be revised. Until then, the conditional acceptance with a request for variance estimates and multi-split validation remains the appropriate verdict.","tokens_in":6344,"tokens_out":8003,"duration_ms":77561,"concrete_test":"Run leave-one-subject-out cross-validation over all 9 subjects using the same three autoencoders, and report per-sequence AUC and classification error for each fold for (a) weighted sum, (b) non-weighted sum, and (c) one-class SVM. If the weighted-sum model does not beat the best baseline by more than the across-subject standard error, the 'best ability' claim in Section 3.3 should be withdrawn or softened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Section 3.3, that the weighted-sum index 'gave the best ability of classification,' is not supported by the reported evaluation. Table 2 lists per-sequence classification error 0.139 for both 'Ours (weighted sum)' and the one-class SVM baseline, so the paper's own comparison is a tie, not a win. More fundamentally, the test set contains 4 subjects, each contributing 9 sequences (1 normal and 8 simulated abnormalities). The 36 per-sequence scores are therefore not independent; the effective sample size for a subject-generalization claim is 4, not 36. No confidence intervals, bootstrap estimates, or cross-validation are reported. At this resolution, a single sequence reclassified changes the error rate by about 1/36 ≈ 0.028, so the differences among per-sequence rows (0.139 vs 0.188 vs 0.250 vs 0.278) are smaller than the granularity of the data. The reported AUC of 0.945 is likewise computed from a single split and could shift substantially with a different subject partition. Because the 'best ability' claim rests on these fragile per-sequence numbers, it is not established by the evidence as presented. The per-frame normalization assumption identified by the reader is also relevant, but even if that assumption is correct, the current split-level statistics are too weak to demonstrate that the weighted index outperforms the simpler baselines.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised gait abnormality index from Kinect skeleton data. Each skeleton is preprocessed by discarding 8 joints and min-max normalizing coordinates per axis; three sparse deep autoencoders are trained on normal walking skeletons only, one per coordinate axis. The abnormality index is a weighted sum of per-axis reconstruction errors (Eq. 2), and per-segment/per-sequence indices are obtained by averaging per-frame scores. The authors evaluate classification of normal vs. eight simulated abnormal gaits from 9 volunteers, comparing with a HMM baseline and a one-class SVM baseline. They report AUC and EER values, with the per-sequence weighted-sum index reaching AUC 0.945 and EER 0.139, and conclude that this weighted-sum index gives the best classification ability.","tokens_in":6616,"tokens_out":6720,"duration_ms":67261,"significance":"If the result holds, the paper offers a simple, unsupervised alternative to hand-crafted gait features, and the per-axis decomposition with sparsity constraints could aid interpretability. Strengths include a one-class formulation that matches practical settings where abnormal gaits are rare and diverse, and explicit comparison with two reimplemented baselines on a common subject split. However, the evidence supporting the central claim is currently thin: all evaluations rely on a single split with four test subjects, and the headline comparison with the one-class SVM baseline is a tie. The significance of the work would be substantially higher if the evaluation uncertainty were quantified and the claimed superiority were supported by statistically meaningful comparisons.","major_comments":[{"comment":"In Table 2, the per-sequence classification error of 'Ours (weighted sum)' is 0.139, identical to that of the One-class SVM [1]. The sentence in Section 3.3 that 'our index estimation that used a weighted sum gave the best ability of classification' is therefore not supported by the paper's own numbers; the method ties the baseline on this metric. Please provide error bars, a statistical comparison, or a clearly qualified claim.","section":"Section 3.3, Table 2"},{"comment":"The evaluation uses a single subject-wise split (5 training, 4 test subjects) and reports point estimates only. The 36 per-sequence scores are not independent because they come from 4 subjects, so the effective sample size for subject generalization is 4, not 36. A single reclassified sequence changes the error rate by about 0.028, which is comparable to or larger than the differences among per-sequence rows in Table 2. I request confidence intervals, leave-one-subject-out cross-validation, or bootstrapping over subjects before the reported AUC/EER can be interpreted as evidence of superiority.","section":"Sections 3.1 and 3.2, Tables 1 and 2"},{"comment":"The method rests on the assumption that gait abnormality is visible in per-skeleton joint positions after per-axis min-max normalization of each skeleton. The paper states this assumption but does not test it. If abnormality is largely a temporal dynamics phenomenon or is related to absolute body or joint positions, this preprocessing could remove the signal. An ablation comparing raw or body-size-normalized coordinates, or a quantitative test of the per-frame assumption, would strengthen the claim.","section":"Sections 2.1 and 2.3"},{"comment":"Several load-bearing hyperparameters are fixed without sensitivity analysis: the sparsity target rho=0.05, the segment length of 20, and the KL/L2 penalty coefficients are not varied, and the optimization details are not reported. In addition, Eq. (2) is typeset ambiguously: as written, w_k = e^{-1}_k sum_k e_k does not define the normalized weights described in the text. Please clarify the formula and report the chosen weighting scheme.","section":"Section 2.2 and Eq. (2)"}],"minor_comments":[{"comment":"The paper says tanh is assigned to two-thirds of the layers, but it does not specify which of the hidden layers use tanh and which use sigmoid; please state the architecture precisely.","section":"Section 2.2"},{"comment":"The batch-loss curves would be much easier to interpret with labeled axes and a legend; without them, the claim that the Z-model loss is less stable is difficult to verify from the figure.","section":"Figure 3"},{"comment":"The HMM row has no per-frame entry; please indicate why this value is unavailable or state explicitly that the method requires a sequence input.","section":"Table 2"},{"comment":"The reimplementations of [1] and [3] are not described in enough detail to be reproduced; please provide the number of HMM states, k-means parameters, one-class SVM kernel and hyperparameters, and any tuning procedure.","section":"Section 3.3"},{"comment":"The final paragraph mentions Procrustes analysis as future work, but since coordinate normalization is central to the proposed preprocessing, a sensitivity experiment with an alternative normalization would be more informative than a future-work note.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The core idea is not circular: the index is reconstruction error from autoencoders trained only on normal data, which is standard unsupervised anomaly detection. The main barrier to acceptance is not methodology but evidence: a single split of nine volunteers cannot support the 'best ability' claim, especially when the per-sequence error ties the one-class SVM baseline. I also note that this is a published ICCE 2018 conference paper; the authors should clarify what is new in this version and whether the journal version includes additional experiments beyond the conference paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, quick read of Nguyen et al. The thing to know: this is a straightforward, honest application of sparse autoencoders to skeleton-based gait scoring. No hand-crafted features; three per-axis autoencoders trained only on normal skeletons; the abnormality index is a weighted sum of reconstruction errors, with weights derived from training MSE. That specific combination is not in the HMM or SVM baselines they cite, and the paper is clear that the setting is unsupervised.\n\nCredit where due: the method is simple enough to reproduce from the text, the per-axis decomposition is sensible, and the hidden-unit visualizations give a nice interpretability check. They also reimplemented two baselines and evaluated on the same train/test split, which is more than many short papers bother to do.\n\nThe soft spots are real and concentrated in the evaluation. The test set has four subjects, each contributing nine gait sequences, so the 36 per-sequence scores are not independent; the effective sample size for subject generalization is four. There are no confidence intervals, no bootstrap, no cross-validation, and no sensitivity analysis for rho=0.05 or the segment length of 20. Table 2 shows the weighted-sum method at 0.139 per-sequence classification error, exactly tied with the one-class SVM baseline. So the Section 3.3 claim that the weighted sum 'gave the best ability of classification' is not established by the paper's own numbers — it is a tie, not a win. Since one reclassified sequence changes the rate by about 0.028, the differences between most rows are within the granularity of the test set. The AUC of 0.945 is also a single-split point estimate and could shift with a different subject partition.\n\nThere is also an explicit assumption, stated in Section 2.1 and echoed in the introduction, that abnormality is visible in per-frame normalized joint coordinates. That assumption is reasonable but untested; if gait abnormality lives mainly in temporal dynamics or body-size differences, the normalization could bury the signal. This is a legitimate concern, not a fatal one.\n\nOverall: the core method is sound as a baseline approach and the writing is honest, but the evidence does not support superiority over simpler alternatives. The paper would need variance estimates, a plain-autoencoder ablation, hyperparameter sensitivity, and ideally released code/data to justify its stronger claims.\n\nWho is it for? Researchers working on markerless gait assessment or reconstruction-error anomaly detection on Kinect skeletons. It is useful as a clean baseline and a clear index formula, not as a major advance.\n\nRecommendation: if this is in the review pipeline, send it to peer review — the idea is legitimate and deserves referees, though they should demand substantial statistical revisions. If it is already published, treat it as a useful reference but do not cite the 'best ability' claim without independent validation.","headline":"A clean, honest per-axis sparse-autoencoder gait index, but the 'best ability' claim rests on a four-subject split and a tie with the one-class SVM baseline.","tokens_in":7137,"tokens_out":2312,"would_cite":false,"duration_ms":24434,"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":"A gait abnormality index can be derived from single skeleton postures alone: sparse deep autoencoders trained only on normal gait score each frame by reconstruction error, and a per-axis weighted, sequence-averaged index separates normal…","keywords":["gait abnormality index","skeleton-based gait analysis","sparse deep autoencoder","reconstruction error","one-class classification","unsupervised learning","depth camera"],"falsifier":"Train the same three autoencoders on normal skeletons only, then test on gait sequences in which the pose sequence is normal but the temporal pattern is altered, for example identical poses with abnormal cadence, hesitation, or phase asymmetry, or the same subjects walking at a different treadmill speed. If the per-sequence weighted index does not separate these from normal walks (AUC near 0.5), the central claim that abnormality is contained in individual postures would be refuted.","tokens_in":6094,"feed_emoji":"🚶","tokens_out":7578,"duration_ms":66459,"temperature":0.7,"pith_summary":"This paper claims that a gait abnormality index can be computed directly from single skeleton postures, without hand-crafted features: three sparse deep autoencoders are trained on normal walking skeletons only, one per coordinate axis, and each new skeleton's reconstruction error becomes an abnormality score. The score for a walking sequence is a weighted average of per-frame errors, with weights set by each axis model's training error. On a dataset of nine volunteers walking on a treadmill, the per-sequence weighted index separates normal from abnormal gaits with AUC 0.945 and equal error rate 0.139, matching the best silhouette-based comparison method while using only skeletal data. If right, this would make automated gait screening unsupervised, inexpensive, and applicable to abnormal gait types never seen during training.","feed_headline":"Skeleton-only autoencoders flag abnormal gait at AUC 0.945","feed_subtitle":"Three sparse autoencoders score single postures; averaging over a walking sequence reaches EER 0.139.","key_machinery":"The load-bearing object is a trio of fully connected sparse autoencoders with a 17-128-32-8-32-128-17 structure, one for each coordinate axis, trained by backpropagation on normal skeletons with a loss combining mean-square reconstruction error, a Kullback-Leibler sparsity penalty with target activation $\\rho=0.05$ on the second layer, and $\\ell^2$ weight decay. The mechanism is one-class reconstruction: an autoencoder trained only on normal postures reconstructs them well and abnormal postures poorly, so reconstruction error translates directly into an abnormality index. The index is the weighted sum of the three per-axis errors with weight $w_k = e^{-1}_k / \\sum_k e_k$, which emphasizes axis models with lower training error; temporal context enters only as post-processing, by averaging the frame indices over a segment of 20 frames or the whole 1200-frame sequence.","core_discovery":"The paper establishes that the normality of a walking posture is largely contained in the single skeleton itself: after discarding eight joints and normalizing each coordinate axis separately to $[0,1]$, a skeleton is fed to three axis-wise deep autoencoders, and the mean squared reconstruction error is a usable per-frame abnormality index. A weighted sum of the three per-axis errors, with weights inversely related to each model's training-set error, outperforms the plain sum, and averaging the per-frame indices over a segment or the whole sequence reduces noise. The central experimental claim is that this unsupervised, feature-free pipeline outperforms a reimplemented hand-crafted skeleton-based HMM baseline and matches a reimplemented silhouette-based one-class SVM baseline, reaching a per-sequence classification error of 0.139.","pith_inferences":["Because normalization scales each skeleton's per-axis coordinates to [0,1], the index is invariant to uniform body-size rescaling; a natural extension would be to test whether the same threshold holds across subjects with very different limb lengths, which the small nine-subject dataset cannot establish.","The paper's low weight on the Z-axis suggests the method depends on viewing geometry: with a side view, where most gait motion is along the optical axis, the axis weighting learned for a frontal view would likely need re-estimation.","The frame-wise scores could be fed into a temporal model such as an HMM or recurrent network instead of averaged, which might capture cadence- or coordination-based abnormalities that per-frame reconstruction error alone misses.","Thresholding individual sparse hidden-unit activations could yield a per-joint abnormality map, a clinically useful extension the paper does not pursue."],"forward_implications":["Screening for gait abnormalities no longer requires labeled abnormal examples: only normal walking skeletons are needed for training, and any deviation that changes joint coordinates becomes detectable as higher reconstruction error.","Because the index is computed per frame, the method can run online: a stream of skeleton frames yields a continually updated abnormality score, with temporal averaging used only to suppress noise.","The per-axis weighted fusion gives an explicit role to each coordinate direction; in a frontal Kinect view, sideways (X) and vertical (Y) joint motion carry more abnormality information than depth (Z).","The sparse middle-layer units act as interpretable joint-difference filters, so the same architecture that produces the index also yields a visualization of which joints and axes drive the abnormality signal.","For gait-type classification, the per-sequence weighted index (AUC 0.945, EER 0.139) supersedes the plain index sum and both reimplemented baselines on the reported dataset."],"supporting_citations":[{"why":"Silhouette-based automatic detection of abnormal gait; provides the one-class SVM baseline the paper reimplements and matches.","marker":"[1]"},{"why":"The authors' earlier skeleton-based HMM with hand-crafted lower-body features; the hand-crafted baseline the autoencoder approach is compared against.","marker":"[3]"},{"why":"Real-time human pose recognition from depth images; supplies the Kinect pose pipeline referenced for skeleton extraction.","marker":"[5]"},{"why":"Efficient human pose estimation from single depth images; the Kinect 2 joint data source that produces the 25-joint skeletons used after joint selection.","marker":"[6]"}],"fun_headline_variants":["Sparse autoencoder flags gait from skeleton alone","Unspervised autoencoder scores gait abnormality","Skeleton-only autoencoder hits EER 0.139","No hand-crafted features: sparse autoencoder detects gait","Single skeleton enough for gait abnormality index"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The key assumption is that abnormal gait leaves a visible mark inside a single normalized skeleton pose, so that after per-axis min-max normalization a normal-trained autoencoder's per-frame reconstruction error separates normal from abnormal movement; if the abnormality lives mainly in timing, coordination, speed, or absolute body geometry, the normalization and frame-wise scoring erase it.","fun_headline_variants_meta":{"raw":{"variants":["Sparse autoencoder flags gait from skeleton alone","Unspervised autoencoder scores gait abnormality","Skeleton-only autoencoder hits EER 0.139","No hand-crafted features: sparse autoencoder detects gait","Single skeleton enough for gait abnormality index"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1288,"prompt_tokens":791,"completion_tokens":497,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":424}},"tokens_in":407,"tokens_out":497,"duration_ms":6224,"temperature":1.0,"reasoning_tokens":424,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:48:13.991840+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same three autoencoders on normal skeletons only, then test on gait sequences in which the pose sequence is normal but the temporal pattern is altered, for example identical poses with abnormal cadence, hesitation, or phase asymmetry, or the same subjects walking at a different treadmill speed. If the per-sequence weighted index does not separate these from normal walks (AUC near 0.5), the central claim that abnormality is contained in individual postures would be refuted.","supporting_citations":[{"cited_title":"Automatic detection of abnormal gait","cited_arxiv_id":null,"evidence_quote":"Silhouette-based automatic detection of abnormal gait; provides the one-class SVM baseline the paper reimplements and matches."},{"cited_title":"Shotton, A","cited_arxiv_id":null,"evidence_quote":"Real-time human pose recognition from depth images; supplies the Kinect pose pipeline referenced for skeleton extraction."},{"cited_title":"Shotton, R","cited_arxiv_id":null,"evidence_quote":"Efficient human pose estimation from single depth images; the Kinect 2 joint data source that produces the 25-joint skeletons used after joint selection."}],"review_version":1}