acceptodds
Under review as a conference paper at ICLR 2027

GaitReader: From Large-Scale Healthy Gait Pretraining to Cross-Center Disease Recognition

Abstract

Six-degree-of-freedom (6-DoF) knee kinematics provide a non-invasive means of assessing lower-limb function and recognizing musculoskeletal disorders. However, existing learning-based approaches largely depend on disease-specific annotations, which are difficult to obtain at scale across diverse patient populations and clinical centers. Although learning from healthy movement has shown promise for gait analysis, whether healthy-only pretraining can support label-efficient knee disease recognition across centers remains insufficiently explored. To investigate this question, we introduce KineGait, a multicenter dataset containing 4,168 healthy participants for self-supervised training and validation, alongside Healthy, anterior cruciate ligament deficiency (ACLD), and knee osteoarthritis (KOA) cohorts for downstream evaluation. Building on this resource, we propose GaitReader, which organizes recordings into cycle-level gait units and learns a gait vocabulary comprising six DoF-specific codebooks. The frozen tokenizer provides discrete targets for contextual pretraining, combining masked code prediction with attribute prediction and waveform reconstruction without disease annotations. Following supervised fine-tuning, GaitReader achieves mean accuracies of 93.04% and 89.87% on fixed internal and external test sets, respectively, across five training–validation folds. Pretraining improves accuracy over the same architecture trained from scratch at every evaluated label fraction; using half of the training labels, it exceeds scratch training with all labels on both test sets. These findings support healthy gait as a source of transferable representations for label-efficient disease recognition across the evaluated centers.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Self-Supervised Learning, Time-Series Representation Learning, Gait Analysis, Six-Degree-of-Freedom Knee Kinematics

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_eTXdOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces KineGait, a multicenter dataset of 6-DoF knee kinematics, including 4,168 healthy participants for self-supervised pretraining and additional cohorts for disease classification. The authors also propose GaitReader, which segments gait recordings into individual cycles, learns DoF-specific gait vocabularies, and uses masked prediction and waveform reconstruction to learn representations from healthy subjects. The pretrained model is then fine-tuned to classify Healthy, ACLD, and KOA subjects. Experiments on internal and external test sets show improved classification performance over several baselines, particularly when fewer labeled training samples are available.

Critical strengths

The paper addresses a meaningful problem: learning useful gait representations from healthy subjects to reduce the need for labeled patient data. The KineGait dataset is a valuable contribution, with over 4,000 healthy participants and additional clinical data collected across multiple centers. The experiments are also fairly comprehensive, including comparisons with several existing methods, external testing, label-efficiency experiments, and ablation studies. The results show consistent improvements from pretraining, especially when fewer labeled samples are available.

Critical weaknesses
  1. Potential confounding in disease classification: Table B.3 shows substantial age and sex differences among the Healthy, ACLD, and KOA groups, while Table B.1 indicates that healthy controls were recruited from community sources and patients from hospitals. Since these factors can also affect gait patterns, it is unclear how much of the classification performance reflects disease-related features rather than differences between the cohorts. I think an age- and sex-matched or stratified evaluation would make the results more convincing.
  2. The contribution of the proposed pretraining method is not fully isolated: Although Table 1 shows good performance, most baselines use different input representations and do not have the same rhythm information as GaitReader. More importantly, removing the gait vocabulary reduces external accuracy only from 89.87% to 88.10% (Table 3), while removing the rhythm embedding causes a much larger drop to 68.10% (Table D.4). This raises the question of how much of the improvement comes from the proposed VQ-based pretraining rather than the cycle-level representation and rhythm information. A more controlled comparison would help clarify this.
Questions for authors
  1. Given the age, sex, and recruitment differences shown in Tables B.1 and B.3, could you provide an age- and sex-matched or stratified evaluation to better understand the impact of these factors on disease classification?
  2. Table D.4 shows a large performance drop when rhythm embeddings are removed. Could you explain why rhythm information has such a strong effect on external performance? Also, would it be possible to compare GaitReader with the strongest baselines using the same cycle-level inputs and rhythm information?
4 comments
AuthorsOpenReview ↗

Response to Reviewer eTXd (Part 1/2)

We thank the reviewer for the constructive review and for highlighting the need to examine demographic differences, explain the contribution of rhythm information to external performance, and strengthen the baseline comparisons. We provide additional analyses and controlled experiments addressing these points, and respond point by point below.

Q1: GaitReader maintains high classification performance in age- and sex-matched cohorts. We evaluate same-sex Healthy/ACLD/KOA triplets with maximum age differences of five and ten years, using the same five fold-specific Full models. Each matched cohort has equal class counts.

Values are five-model means (%).

Test cohortSubjectsAccuracyMacro-F1
Original internal13893.0492.09
Matched internal (5 years)991.1190.86
Matched internal (10 years)1890.0089.98
Original external31689.8789.42
Matched external (5 years)4587.5687.35
Matched external (10 years)7289.7289.66

The matched external accuracies remain 87.56% and 89.72%, compared with 35.56% and 39.17% for age-and-sex-only classifiers on the same cohorts. These results support the predictive value of gait representations beyond those demographic features.

Age differences also decrease under both matching settings. The pairwise absolute age SMDs are shown below; sex proportions are identical across the three classes in every matched cohort.

TestPopulationHealthy–ACLDHealthy–KOAACLD–KOA
InternalOriginal2.8190.4272.817
InternalMatched (5 years)0.1910.0290.191
InternalMatched (10 years)0.3660.2200.519
ExternalOriginal1.7380.6372.805
ExternalMatched (5 years)0.0750.0420.038
ExternalMatched (10 years)0.1680.1640.395

SMD uses the original-cohort pooled SD for each class pair as a fixed denominator across the original and both matched settings within each test set. The five-year setting yields closer age balance; the ten-year setting retains more subjects.

Revised manuscript: Appendix D.7 / Table D.10 (p. 31) for protocol and performance, and Table D.11 (p. 32) for demographic composition and age SMDs.

AuthorsOpenReview ↗

Response to Reviewer eTXd (Part 2/2)

Q2: Removing rhythm mainly impairs KOA and Healthy recognition, with substantially larger recall decreases externally. ACLD recall changes little.

Recall values are means across five fold-specific models (%). Duration is the mean of subject-level cycle means (seconds).

TestClassFull recallw/o rhythm recallMean cycle duration
InternalHealthy94.8087.600.876
InternalACLD92.5092.191.411
InternalKOA90.8376.671.796
ExternalHealthy85.2669.050.946
ExternalACLD85.2483.501.197
ExternalKOA97.6353.902.314

The duration statistics help explain this pattern: KOA cycles are longer than Healthy cycles, with a larger difference externally (2.314 vs 0.946 s) than internally (1.796 vs 0.876 s). The longer cycles are consistent with reduced walking speed and cadence and increased stride time reported for KOA by Gök et al. (2002). Resampling cycles to 100 points removes their absolute duration; duration/interval inputs restore this information alongside morphology. Together, these observations help explain the larger external removal effect and support the value of rhythm information across the evaluated center populations.

The gain also depends on pretraining, not just rhythm inputs. With the same backbone and rhythm inputs, the following five-model mean external accuracies are obtained:

ConfigurationExternal Acc. (%)
Supervised backbone82.03
Shape–attribute coupled reconstruction83.61
Full GaitReader89.87

We further tested HeartLang and TS2Vec, the strongest internal and external baselines in Table 1, respectively. HeartLang already uses the same gait-cycle units as GaitReader; we add rhythm inputs and also test greater depth and bilateral cross-attention. For TS2Vec, we compare continuous and cycle-aligned inputs, each with and without rhythm. Values below are five-model mean accuracy (%); “Deeper” denotes four contextual layers.

Model / variantInternal Acc.External Acc.
HeartLang: Original91.4574.56
HeartLang: Original + rhythm91.0179.81
HeartLang: Deeper90.2978.35
HeartLang: Deeper + rhythm90.1480.38
HeartLang: Deeper + rhythm + cross90.7283.73
TS2Vec: Original87.2585.76
TS2Vec: Original + rhythm89.2885.57
TS2Vec: Original + cycle86.5270.06
TS2Vec: Original + cycle + rhythm89.1381.90
Full GaitReader93.0489.87

GaitReader retains its advantage on both test sets after the baselines receive cycle-level inputs and rhythm information. These comparisons support contributions from the learned gait representations alongside explicit timing.

Revised manuscript: class-wise recall and duration analysis in Appendix D.8, Table D.12 and Figure D.2 (p. 33); HeartLang/TS2Vec comparisons in Appendix D.6, Tables D.7–D.8 (p. 29); same-backbone pretraining comparisons in Table D.9 (p. 29; discussion on p. 30).

Reviewer_eTXdOpenReview ↗

Thank the authors for their rebuttal and experiments

Thank the authors for their rebuttal and experiments.

For my first concern, I think it's good they have an additional dataset to verify the issue, but the number of new samples are still limited thus the evaluation result is still not convincing enough. And the issue of data source still exists.

My second concern about the rhythm embedding is addressed with their new experiments, I thus raise my ratings to 3.

AuthorsOpenReview ↗

Response to Reviewer eTXd

Thank you for your thoughtful follow-up. We appreciate your careful review and constructive feedback, which have helped us improve the paper and strengthen its evaluation.

Reviewer_xw9fOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces KineGait and GaitReader, using healthy-gait vocabulary learning and masked contextual pretraining for Healthy/ACLD/KOA classification. It achieves 93.04% internal and 89.87% external accuracy. Same-architecture comparisons demonstrate improved training-label efficiency through healthy-only pretraining.

Critical strengths

1.Same-architecture comparisons support pretraining. The pretrained model using half the training labels outperforms scratch training with all labels, demonstrating benefits beyond architectural differences.

2.External evaluation and detailed reporting strengthen the evidence. Subject-level partitioning, a fixed external test cohort, and extensive implementation details improve experimental credibility and auditability.

Critical weaknesses

1.Diagnosis is confounded with demographics and recruitment source. Mean ages in the Healthy, ACLD, and KOA training/validation groups are 58.52, 30.60, and 64.02 years, respectively, with substantial sex differences. Controls are community-recruited, whereas patients are hospital-recruited. Similar structure persists externally, so cross-center performance does not rule out demographic or recruitment-related cues. Disease-specific discrimination remains insufficiently established.

2.The vocabulary’s contribution and component interactions remain unclear. External accuracy is 89.87% for the full model, 88.10% without the vocabulary, and 68.10% without rhythm embeddings. Existing ablations mainly remove components from the full model, and some change multiple supervisory components simultaneously. A progressive evaluation starting with cycle segmentation and supervised training of the same backbone, then adding rhythm information, direct reconstruction pretraining, and vocabulary supervision, is missing. The incremental value of the complex pretraining design over simpler configurations therefore remains unresolved.

3.The architecture figure needs clearer academic presentation. Figure 1 contains excessive colors and nested boxes, small labels, and an unclear visual hierarchy, making the training stages, frozen components, and supervision paths difficult to follow. Simplifying visual elements, standardizing typography and module styles, and distinguishing data flow from supervision would improve readability. This is a presentation concern rather than an independent reason for rejection.

Questions for authors

1.Do the main findings persist after age/sex matching or reweighting? How well does a demographics-only classifier perform?

2.How does the same backbone perform with cycle segmentation and supervised training alone? What are the incremental gains from adding rhythm information, direct reconstruction pretraining, and vocabulary supervision?

3.With identical inputs, backbone, and attribute supervision, how much does the vocabulary approach improve over direct reconstruction? How competitive is classification using VQ-Gait features directly?

5 comments
AuthorsOpenReview ↗

Response to Reviewer xw9f (Part 1/2)

We thank the reviewer for the careful review and for raising important questions about demographic differences and the contributions of healthy-gait pretraining and vocabulary-guided learning. We address these questions through demographic matching and controlled pretraining comparisons, and respond point by point below.

Q1: GaitReader retains strong performance after age/sex matching and outperforms demographics-only classifiers. We form same-sex Healthy/ACLD/KOA triplets with a maximum age difference of five or ten years, giving equal class counts in each matched cohort. Both settings evaluate the same five fold-specific models.

Entries are mean accuracy/macro-F1 (%); n is the number of subjects.

Test cohortnFull GaitReaderAge + sexAge + sex + BMI
Original internal13893.04 / 92.0967.83 / 62.5268.99 / 64.12
Matched internal (5 years)991.11 / 90.8633.33 / 16.6733.33 / 16.67
Matched internal (10 years)1890.00 / 89.9834.44 / 28.3341.11 / 35.17
Original external31689.87 / 89.4264.87 / 62.3662.41 / 60.54
Matched external (5 years)4587.56 / 87.3535.56 / 32.3532.00 / 28.73
Matched external (10 years)7289.72 / 89.6639.17 / 37.8136.39 / 35.17

High classification performance is retained under both matching settings, while age, sex, and BMI alone give substantially lower performance. This supports predictive information in gait representations beyond the evaluated demographic features.

Matching reduces every pairwise age difference, with closer balance under the five-year limit. The absolute age SMDs are:

TestPopulationHealthy–ACLDHealthy–KOAACLD–KOA
InternalOriginal2.8190.4272.817
InternalMatched (5 years)0.1910.0290.191
InternalMatched (10 years)0.3660.2200.519
ExternalOriginal1.7380.6372.805
ExternalMatched (5 years)0.0750.0420.038
ExternalMatched (10 years)0.1680.1640.395

For each class pair, SMD divides the absolute age-mean difference by the original-cohort pooled SD, fixed before and after matching within each test set. Sex proportions are identical across the three classes in each matched cohort.

Revised manuscript: matching protocol, SMD definition, and performance in Appendix D.7 / Table D.10 (p. 31); demographic composition and age SMDs in Table D.11 (p. 32).

AuthorsOpenReview ↗

Response to Reviewer xw9f (Part 2/2)

Q2: The requested comparison shows successive gains from rhythm inputs, healthy-gait pretraining, and the full vocabulary-guided objectives. All configurations use the same backbone, subject folds, and fixed test sets.

Values are five-model means (%).

ConfigurationInternal Acc.Internal Macro-F1External Acc.External Macro-F1
Supervised backbone, w/o rhythm86.6784.6572.9772.15
Supervised backbone, with rhythm88.4187.4982.0381.51
Direct waveform pretraining90.2989.3588.1087.80
Full GaitReader93.0492.0989.8789.42

Full GaitReader performs best on all four metrics. Rhythm improves the supervised baseline; healthy pretraining improves it further; the complete vocabulary-guided objectives outperform direct waveform pretraining.

Revised manuscript: Appendix D.6, Table D.9 (p. 29), with discussion continuing on p. 30.

Q3: GaitReader outperforms both attribute-coupled direct reconstruction and direct VQ-Gait feature classification. We added the following controls. Values are five-model means (%).

ConfigurationInternal Acc.Internal Macro-F1External Acc.External Macro-F1
Shape–attribute coupled reconstruction90.0088.9083.6183.76
VQ-Gait features, linear probe84.3582.7384.6883.97
VQ-Gait features, fine-tuning90.0088.1577.8578.41
Full GaitReader93.0492.0989.8789.42

The reconstruction control retains attribute supervision and its role in waveform recovery. It predicts continuous shape, mean, and scale, then reconstructs the waveform from those predictions. Waveform loss updates both shape and attribute heads. With the same inputs, backbone, masking rules, and attribute supervision, Full improves external accuracy and macro-F1 by 6.27 and 5.66 percentage points, respectively.

For direct VQ-Gait classification, we pool pre-quantization features and continuous attributes over cycles within each side and DoF, concatenate them, and train a linear classifier. Linear probing freezes the VQ encoder; fine-tuning updates it. Both are outperformed by Full on all four metrics, supporting the use of the healthy vocabulary to pretrain contextual GaitFormer rather than classifying pooled local features alone.

Revised manuscript: Appendix D.6, Table D.9 (p. 29), with discussion on p. 30.

AuthorsOpenReview ↗

Response to Reviewer xw9f

Thank you again for your careful review. We would appreciate your feedback on whether our rebuttal and additional experiments have addressed your concerns. If you have any remaining questions or would like further clarification, we would be happy to continue the discussion.

Reviewer_1N1TOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The work proposes GaitReader, a self-supervised method for recognizing knee disorders from 6-DoF knee kinematics. It also introduces KineGait, that is a multicenter dataset covering 4168 healthy subjects that is used for pretraining and it also has subjects having anterior cruciate ligament deficiency (ACLD), knee osteoarthritis (KOA) that are used for downstream evaluation. The proposed GaitReader first uses to segment recordings into individual gait cycles, then it uses VQ-Gait that learns six DoF-specific codebooks that they represent recurring waveform patterns as discrete tokens. Finally, they propose GaitFormer, a transformer-based encoder, that is pretrained on healthy gait data using (i) masked code prediction, (ii) attribute prediction, and (iii) waveform reconstruction objectives. Once pretrained, they finetune it for disease classification. Experiments show that GaitReader outperforms both supervised and self-supervised approaches and achieves 93.04% accuracy on the internal test set and 89.87% on the external test set.

Critical strengths
  • The proposed KineGait dataset has 4168 healthy subjects that can be used for self-supervised pretraining and it also includes labeled cohorts that is useful for downstream disease classification. The fact that it is large scale and also is multicentered gives a valuable contribution for the research community.
  • The work shows a clear motivation that scaling disease-specific labeled data is hard to scale but we can learn useful representations only from healthy people yet transfer it effectively to disease classification and reduce this depdence of large-scaled labeled pathology data. The experiments also show that even when using 50% of training labels, it can suppress baselines that used 100% of the labels and shows a clear strength of pretraining.
  • The work has different design choices and also provides good ablations to investigate the effect of individual pretraining objectives, gait-cycle segmentation, bilateral attention, input embedding, vocabulary size, and masked ratios. These should help future research to get insight of what design choices are important and how effective they are.
  • Unlike recent approaches that might be bounded by internal evaluation, the work also tests on independent hospital cohort and shows it can transfer to the evaluated clinical centers.
Critical weaknesses
  • The work largely combines existing previously used approaches like gait-cycle segmentation, vector quantization, masked token prediction, and transformer-based contextual modeling. Several components, like shape-attribute disentanglement and vector quantization of waveform shapes, are similar to VQShape. Also, two-stage pretraining, masked discrete-token prediction, and the use of a physiological vocabulary are similar to HeartLang. While the combination of these approaches is shown to be effective, the work does not propose a fundamentally new approach to learning representations.
Questions for authors
  • What are the key methodological contributions that distinguish GaitReader from these approaches?
  • What motivates using six separate DoF-specific codebooks rather than a shared codebook across all motion channels?
  • Is the improvement over HeartLang mainly comes from the proposed learning objectives, the rhythm embeddings, the bilateral attention mechanism, or differences in architectural capacity?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 1N1T (Part 1/2)

We thank the reviewer for the thoughtful review and for highlighting the importance of clarifying our methodological contributions, the rationale for DoF-specific codebooks, and the comparisons with HeartLang. We address these points with methodological explanations and additional controlled experiments, and respond point by point below.

Q1: GaitReader changes the movement-unit construction, the representation being quantized, and the way the vocabulary supervises the contextual backbone. We build on shape–attribute separation from VQShape and two-stage vocabulary-based pretraining from HeartLang, with the following distinctions.

DesignVQShapeHeartLangGaitReader
Representation unitSubsequences with learned locations and lengthsQRS-centered heartbeat unitsComplete gait cycles with shared boundaries across the six DoFs within each side
Input to quantizationShape latents derived from contextual sequence representationsContextual heartbeat embeddingsIndependently encoded, normalized cycle shapes and DoF-specific codebooks
Role of the vocabularyShape–attribute reconstruction for learning token representationsTargets for masked code-ID predictionFrozen shape templates coupling code prediction, attribute prediction, and masked waveform recovery

GaitParser makes the representation cycle-structured. It uses flexion–extension kinematic anchors to define movement units, preserves within-cycle correspondence across DoFs, and retains duration and inter-cycle interval after resampling. GaitFormer models these units through within-side multi-DoF attention and bilateral cross-attention, without requiring cycle-to-cycle alignment between independently recorded limbs.

VQ-Gait separates local shape learning from contextual modeling. Each normalized cycle is encoded independently by a DoF-specific CNN and quantized in its corresponding codebook. Its measured mean and scale bypass quantization. Unlike quantizing contextual embeddings, this learns local motion-shape templates before GaitFormer learns cross-cycle and bilateral relationships.

The second-stage objectives are coupled through waveform recovery. GaitFormer predicts code probabilities, mean, and log-scale at masked positions. The probabilities weight frozen decoded shape templates; the predicted scale and mean then restore the waveform. Waveform loss therefore supervises both the code and attribute predictions. The distinction from VQShape is the use of a frozen healthy vocabulary to train a separate contextual backbone; the distinction from HeartLang is this joint recovery beyond code-ID prediction.

These designs are described in Sections 3.1–3.4, with component ablations in Section 4.5 and Table 3 (p. 9).

AuthorsOpenReview ↗

Response to Reviewer 1N1T (Part 2/2)

Q2: We use DoF-specific codebooks to learn a dedicated vocabulary of movement patterns for each kinematic component. Similar normalized waveforms in different DoFs can describe different movements—for example, a flexion–extension pattern and an internal–external rotation pattern have different kinematic meanings even when their curve shapes resemble each other. Separate codebooks allow quantization to focus on shape variations within each DoF, rather than requiring all components to share one pool of prototypes. This provides GaitFormer with DoF-specific motion-shape targets for masked prediction, encouraging it to learn how each component’s movement pattern relates to the surrounding gait context. Each codebook is shared across subjects and sides so that the same component is represented consistently across recordings.

We tested this choice against a shared codebook with the same total number of codes and code dimension. Values below are five-model mean accuracy/macro-F1 (%).

Codebook organizationTotal codesCode dimensionInternal Acc. / Macro-F1External Acc. / Macro-F1
Shared: one 768-code codebook76812891.74 / 90.9287.66 / 87.09
DoF-specific: six 128-code codebooks76812893.04 / 92.0989.87 / 89.42

DoF-specific codebooks perform better on both test sets, supporting their use for disease classification.

Revised manuscript: motivation in Section 3.3, opening paragraph (p. 5); results in Section 4.5 and Table 3 (p. 9); comparison definition in Appendix C.5, Table C.3 (p. 24).

Q3: The gains reflect contributions from rhythm inputs, bilateral attention, and the pretraining objectives; increasing model capacity alone does not close the gap. We tested these factors in HeartLang while retaining its tokenizer and discrete-code objective.

Values are five-model mean accuracy (%); Params counts active downstream parameters. “Deeper” increases contextual depth from two to four layers at width 128.

ModelParamsInternal Acc.External Acc.
HeartLang: Original363,39591.4574.56
Original + rhythm380,29191.0179.81
Deeper627,58790.2978.35
Deeper + rhythm644,48390.1480.38
Deeper + rhythm + cross909,95590.7283.73
Full GaitReader732,41993.0489.87

The deeper HeartLang with rhythm and cross-attention has more parameters than GaitReader, yet its external accuracy remains 6.14 percentage points lower.

The same-backbone controls and GaitReader ablations also support the proposed objectives and bilateral modeling:

ConfigurationInternal Acc.External Acc.
Shape–attribute coupled reconstruction90.0083.61
w/o bilateral attention92.1786.96
w/o attribute and waveform losses91.8888.61
w/o attribute loss92.0388.54
w/o waveform loss92.7587.72
Full GaitReader93.0489.87

The coupled-reconstruction control retains the backbone, inputs, masking, attribute supervision, and attribute-mediated waveform recovery, but replaces vocabulary-guided shape prediction with continuous regression. Full retains an advantage with rhythm information already present, and removing bilateral attention or the attribute/waveform objectives also lowers external accuracy.

Revised manuscript: HeartLang controls in Appendix D.6, Table D.7 (p. 29); coupled-reconstruction control in Table D.9 (p. 29); GaitReader ablations in Section 4.5 and Table 3 (p. 9).

I am thankful to the authors for their elaborate rebuttal and extra experiments.

Clarifications about methodological aspects help to differentiate GaitReader from VQShape and HeartLang. For example, the clarification about cycle-based representation and DoF-based quantization, as well as joint training targets, make the proposed approach much more understandable.

Regarding my second question, extra experiments related to the comparison of shared and DoF-based codebooks are quite helpful. Moreover, the extended experiments on HeartLang help to understand that better performance cannot be explained by architectural differences only, rhythm embeddings, or bilateral attention mechanisms.

The only thing that I still find doubtful about the paper is the novelty of the methodology. Even though the proposed combination works well and is quite justified, several elements of the approach are the modifications of the previous works.

However, I like the experimental results, thorough ablations, and multicenter dataset used in the paper. All my questions were properly answered in the rebuttal.

Thus, I keep my score of 3 (Weak Acceptance).

AuthorsOpenReview ↗

Response to Reviewer 1N1T

Thank you for your thoughtful follow-up. We are pleased that our clarifications and additional experiments have addressed your questions and made the methodological distinctions clearer.

We acknowledge that several elements of our approach build on and modify prior work. However, the design and refinement of these elements are guided by our understanding of gait modeling and the gait characteristics that the learned representations should capture. Specifically, cycle-based representation preserves meaningful movement units, DoF-specific codebooks represent distinct kinematic components, and joint shape–attribute recovery captures both movement morphology and continuous variation. These considerations motivate a coherent framework for learning transferable gait representations from healthy populations, supported by controlled experiments and multicenter evaluation.

Thank you again for your careful review and constructive feedback, and for your positive assessment of our experimental results and dataset. Your suggestions have helped us clarify the methodological contributions and strengthen the evaluation in the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.