Abstract

Turning difficulty is an early and clinically significant indicator of Parkinson’s disease (PD). Skeleton-based approaches are already widely used in clinical movement analysis, as they offer a markerless and privacy-preserving representation of human motion from ordinary video. However, automated turning assessment from skeleton sequences remains challenging due to the temporal complexity of turning movements. Existing skeleton-based methods either treat the entire sequence globally, ignoring its inherent temporal structure, or require expensive frame-level phase annotations. We propose a weakly-supervised temporal decomposition framework that discovers discriminative temporal patterns from video-level diagnostic labels alone. The framework integrates a transition detector operating on motion gradients with spatial-temporal graph convolutional networks via learned soft masks, decomposing the turning sequence into multiple temporal instances with diagnostic relevance learned through attention aggregation. A learnable adjacency mechanism further adapts the skeleton graph to emphasize clinically relevant body connections. Decoupling boundary detection from the classification objective allows the learned phase structure to serve as an independent clinical interpretation, unbiased by diagnostic labels. Evaluated on the clinical PD dataset, our method achieves an AUC-ROC of 91.05% under 5-fold subject-level cross-validation, while providing interpretable outputs including instance-level importance weights and anatomical relevance maps consistent with clinical knowledge of Parkinsonian turning symptoms. The code is available at https://github.com/jieming1113/TempoDecompGCN-PD

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/0926_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/jieming1113/TempoDecompGCN-PD

Link to the Dataset(s)

REMAP dataset: https://data.bris.ac.uk/data/dataset/21h9f9e30v9cl2fapjggz4q1x7

BibTex

@InProceedings{ZhaJie_WeaklySupervised_MICCAI2026,
        author = { Zhang, Jieming AND Chung, Tai-Myoung AND Park, Hogun},
        title = { { Weakly-Supervised Temporal Decomposition with Graph Neural Networks for Parkinsonian Turning Assessment } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16880},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    Main contributions are: 1) weakly-supervised temporal decomposition framework 2) a dual-branch architecture 3) evaluation on a clinical dataset

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The main strength of this paper is the proposed architecture, which introduces a novel and well-motivated approach to Parkinsonian turning assessment. The ablation study is clearly demonstrates the contribution of each component. The comparison with existing methods is also convincing, with the proposed method outperforming all baselines across all reported metrics.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    Several comments on the text:

    1) The terms task-specific and phase-specific temporal structure are central to the paper but never explicitly defined. Adding a short definition in the Introduction would help readers coming from a general ML background.

    2) The term ‘phase boundaries’ is used in the Introduction before being formally defined in Section 3.2.I would suggest either adding a brief informal definition at first use in the Introduction, or at minimum a forward reference to Section 3.2.

    3) In the Introduction, two consecutive sentences essentially say the same thing: ‘learning phase boundaries directly from video-level diagnostic labels alone’ and ‘without any frame-level supervision’ convey the same idea. I’d suggest merging these into one sentence to avoid repetition.

    4) The claim that frame-level annotations are impractical in clinical settings is made in Related Work and then repeated again in Section 3.2.

    5) The limitation of fixed skeleton topology (treating all joint connections equally) is only introduced in Section 3.3, but I think it would fit better in the Introduction alongside the global pooling limitation. This way readers understand from the beginning that the paper addresses both temporal and spatial rigidity in existing methods, which better reflects the full scope of the contribution.

    More broadly, some motivational points are repeated across sections in slightly different wording. A careful pass to either consolidate or use forward/backward references would tighten the paper considerably.

    6) What is the difference between the ST-GCN baseline in Table 1 (which achieves 61.71% AUC) and the ST-GCN method in Table 3 (69.94% AUC)? Should they be identical?

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The authors convincingly proved their concept.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    As I stated at the beginning the authors convincingly proved their concept, but they need to reorganise their paper.



Review #2

  • Please describe the contribution of the paper

    From turning sequences of Parkinson videos, the work proposed a weakly-supervised temporal decomposition framework from 2D skeleton sequences and using only diagnostic labels. The method detects soft phase boundaries, and aggregates instance-level features using a gated MIL attention module. Then a standar ST-GCN backbone is adjusted to predict Parkinson from controls in a binary classification.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • a dual-branch architecture that aims to decouple temporal boundary
    • an adaptive adjacency mechanism in ST-GCN
    • The overall pipeline is well structured and diagrammed, clarifying how boundary scores
    • Addresses a clinically important and underexplored problem—turning assessment
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The central idea of decoupling boundary detection from classification is attractive for interpretability, yet the current training of the decomposition branch appears underconstrained.
    • For the sparsity and smoothness there is no clear how link with motion saliency
    • The adaptive adjacency design is sensible and constrained by the physical skeleton adjacency.
    • There is necessary a Statistical robustness study. Statistical test that evidence significant differences
  • Please rate the clarity and organization of this paper

    Satisfactory

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper is well written and methodologically sound. The idea of decoupling detection is attractive but there is not clear how losses are properly implemented to contribute to main goal of the worl Additionally, the authors should justify sensibility of adjacency vs skeletons. Statistical test are demanding to evidence contribution of the work. Also, it is demanding how were diagnosed PD, under which scale. How many patients tthere exist in each progression level. H&Y? UPDRSS?..how was partition regarding such scales.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    This paper addresses automated assessment of turning in Parkinson’s disease (PD) patients using 2D skeleton sequences extracted from video. The core challenge is that turning exhibits an inherent temporal structure that is clinically relevant, yet most existing methods either process the sequence globally or require expensive frame-level annotations of movement phases. The authors propose a weakly-supervised framework that decomposes the sequence into temporal instances via learned boundary scores derived from motion gradients. A diagnostic branch based on ST-GCN with adaptive adjacency and MIL-style gated attention aggregates instance-level representations into a video-level prediction. Evaluated on the REMAP clinical dataset with 1,294 sequences and 5-fold subject-level cross-validation, the method achieves 81.54% accuracy and 91.05% AUC-ROC, outperforming several baselines.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    1.The paper tackles a clinically relevant and well-motivated problem. Turning is a meaningful indicator in PD assessment, and skeleton-based analysis provides a practical and privacy-preserving representation.

    1. The methodological contribution is clear and reasonably novel. The approach goes beyond standard temporal attention by explicitly discovering ordered temporal instances in a weakly-supervised manner and integrating them with GCN-based modeling. The decoupling of boundary learning from classification is conceptually interesting and supports interpretability. 3.The experimental evaluation is solid within scope. The paper includes meaningful ablations, comparisons with relevant baselines, and interpretability analysis. Performance improvements, especially in AUC-ROC, are substantial.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    1.The interpretation of the discovered temporal instances as clinically meaningful phases is somewhat overstated. Boundary detection is driven only by sparsity and smoothness constraints, and classification loss does not influence boundary learning. While this supports interpretability claims, it does not strongly validate that the learned segments correspond to true biomechanical phases. 2.Reproducibility is moderate. While several experimental details are provided, important aspects remain unclear, including some hyperparameters, such as the weighting coefficients (λₛ and λₘ) of the auxiliary loss terms. 3.Evaluation is limited to a single clinical dataset. The authors acknowledge this limitation, but it still restricts the strength of conclusions regarding real-world clinical applicability and generalization.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?

    N/A

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Strengths include a clear methodological contribution and strong experimental performance. Weaknesses include limited validation scope, moderate reproducibility, and somewhat overstated interpretability claims. Overall, the strengths outweigh the weaknesses.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Accept

  • [Post rebuttal] Please justify your final decision from above.

    After reading the rebuttal, I consider that the authors have satisfactorily addressed most of my main concerns. In particular, they clarified that the learned temporal segments should not be interpreted as validated biomechanical phases, but rather as soft temporal transition candidates/instances. This substantially mitigates my concern that the interpretability claims were overstated. The authors also addressed the reproducibility concern by providing the missing auxiliary loss weights, and they clarified the apparent discrepancy between the lightweight ST-GCN ablation baseline and the full ST-GCN architecture. These clarifications are satisfactory.

    The main remaining limitation is the evaluation scope. Although the dataset contains a relatively large number of sequences, the rebuttal clarifies that these correspond to only 12 PD and 12 HC participants. The use of subject-level 5-fold cross-validation is appropriate and helps avoid data leakage, but the small number of subjects and the lack of severity-stratified analysis still limit the strength of the clinical generalization claims.

    Overall, I consider that the rebuttal resolves the main methodological and presentation concerns, while the remaining limitations are acceptable.



Author Feedback

We thank all reviewers for their thorough and constructive feedback. We address the major concerns below. [Q1] Presentation suggestions (R1-W1–W5) All five presentation suggestions will be addressed in the revision. [Q2] Baseline vs full model of ST-GCN: Table 1 and Table 3 (R1-W6) The “ST-GCN baseline” in Table 1 is our lightweight 2-block encoder, serving as the minimal ablation baseline (61.71% AUC). The “ST-GCN [27]” in Table 3 is the full Yan et al. architecture, trained under the same protocol (69.94% AUC). The 8.23% gap reflects this capacity difference, not an inconsistency; a clarifying footnote will be added. [Q3] Training constraints and link between losses and motion saliency (R2-W1,W2) The decomposition branch is architecturally decoupled, but its boundary estimates remain explicitly constrained. Boundary scores s(t) are computed from [f(t);Δf(t)], capturing motor state and motion change, and are constrained by a learnable position prior, temporally ordered K=3 masks, and L_sparse/L_smooth. The auxiliary losses suppress dense or fragmented detections. By design, non-differentiable boundary selection prevents L_cls from directly biasing boundary locations, preserving label independent transition discovery, while the resulting masks still serve as weights in mask-weighted pooling and ultimately determine the sequence representation used for classification. We will clarify this path and describe the learned regions as soft temporal transition candidates, not ground truth phases. The +5.93% AUC gain over fixed decomposition (Table 1, (p<0.01)) supports this design. [Q4] Adaptive adjacency design (R2-W3) Physical skeleton topology serves as an anatomically informed architectural prior, constraining learnable adjacency to anatomically valid connections, limiting parameter space on clinical-scale data and preserving interpretability of learned edge weights. We agree that non-local dynamics carry clinical value; capturing such interactions is a natural direction for future work. [Q5] Statistical robustness (R2-W4) All tables report mean±standard deviation across the same subject-level 5-fold CV. Table 1 already includes paired t-tests (*p<0.01, **p<0.001), confirming each component is statistically significant; the last row of Tables 2–3 is also statistically significant over all other configurations; significance notes will be added in the revision. [Q6] Clinical characterization (R2) Our evaluation uses the REMAP dataset. PD diagnosis followed the UK Brain Bank Criteria. The 1,294 sequences (773 PD, 521 HC) derive from 12 PD and 12 HC participants. H&Y=2.3±0.8 (entry criterion ≤3, mild-to-moderate PD); per-H&Y-stage breakdown is not reported. MDS-UPDRS III=36.8±23.0.A clinical summary will be added to Section 4.1.Subject-level 5-fold CV prevents data leakage; fold assignment was not stratified by severity score, as our framework operates solely on video-level diagnostic label; severity-stratified analysis remains future work. [Q7] Interpretability claims (R3-W1) We agree that the learned segments should not be interpreted as validated biomechanical phases. “Early/Mid/Late-Turn” are temporal-ordering names for soft instances, not frame-level phase annotations. Fig. 2 provides qualitative interpretation only: PD shows higher late-instance attention than HC (α₃=0.51 vs 0.36), and the anatomical map emphasizes lower-limb connections, consistent with prior PD turning/gait findings. We will replace “clinically meaningful phase boundaries” with “soft temporal transition candidates/instances” and state that direct frame-level phase validation remains future work. [Q8] Reproducibility: missing hyperparameters (R3-W2) Missing λ_s=0.05 and λ_m=0.1 will be added to Section 4.1; others are already reported. Code will be released upon acceptance. [Q9] Single-dataset (R3-W3) We acknowledge this limitation. REMAP is among the few public labeled datasets for PD turning assessment, and cross-cohort generalization remains future work.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.

    There are some discrepancies among the three reviewers, consisting of one weak reject, one weak accept, and one accept. The main contributions of the paper include the proposed dual-branch weakly supervised temporal decomposition framework, that is evaluated on a clinical dataset with promising results. Meanwhile, there are a number of aspects of the work that raise concerns. It would be beneficial if the authors provide further input and clarifications in the areas specifically brought out by the reviewers, including but not limited to, how to ensure constrained training process in current system esp. by engaging the current loss functions, empirical studies of the statistical robustness aspect; details of the empirical evaluations; possibly overstated claims; reproducibility; improvement of the current presentation.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    Most of the major concerns have been addressed by the rebuttal. That said, there remain concerns such as the limited number of subjects in dataset. The authors are encouraged to carefully follow the reviews and revise the draft accordingly, if it is eventually accepted by this MICCAI.



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation.

    The paper deals with the automated analysis of the Parkinson’s disease turning assessment, which is a clinically well-motivated problem. Specifically, the authors have proposed a dual-branch architecture based on skeleton-based processing backbones.

    The reviewers identified that this is an important and clinically motivated problem to be solved and appreciate the dual-branch architecture. However, several concerns were raised regarding the choice of the loss functions and how they relate to the main goal of the paper and most importantly the limited size of the dataset especially in terms of the number of subjects (only limited to 12PD and 12 healthy controls), which questions the generalzability and robustness and experimental validation of the claims of the proposed framework. Moreover, I also think that several major baselines in the general skeleton-based human activity understanding domain are missing - for instance HyperGCN (modeling skeleton sequences as a hypergraph). The authors built on the very old baselines of skeleton-based GCN processing where ST-GCN was one of the first baselines, but have not explored how more recent approaches compare to it. Thus, the gains presented by the authors may be due to the inherent incapability of methods like ST-GCN to extract relevant features necessary for this task.



Meta-review #3

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation.

    After rebuttal, two reviewers are satisfied with the rebuttal. The authors have properly addressed most major concerns including interpretability, reproducibility and architectural discrepancy. Yet limited participant count and absence of severity-stratified analysis weaken clinical generalization. Overall, this is a good paper and the recommendation is Accept.



back to top