Abstract

Video-based assessments offer a scalable pathway for remote Parkinson’s disease (PD) screening. While traditional approaches rely on handcrafted features mimicking clinical scales, recent advances in video foundation models (VFMs) enable representation learning without task-specific customization. However, the comparative effectiveness of different VFM architectures across diverse clinical tasks remains poorly understood. We present a large-scale systematic study using a novel video dataset from 1,888 participants (727 with PD), comprising 32,847 videos across 16 standardized clinical tasks. We evaluate seven state-of-the-art VFMs – including VideoPrism, V-JEPA, ViViT, and VideoMAE – to determine their robustness in clinical screening. By evaluating frozen embeddings with a linear classification head, we demonstrate that task saliency is highly model-dependent: VideoPrism excels in capturing visual speech kinematics (no audio) and facial expressivity, while V-JEPA proves superior for upper-limb motor tasks. Notably, TimeSformer remains highly competitive for rhythmic tasks like finger tapping. Our experiments yield AUCs of 76.4–85.3% and accuracies of 71.5–80.6%. While high specificity (up to 90.3%) suggests strong potential for ruling out healthy individuals, the lower sensitivity (43.2–57.3%) highlights the need for task-aware calibration and integration of multiple tasks and modalities. Overall, this work establishes a rigorous baseline for VFM-based PD screening and provides a roadmap for selecting suitable tasks and architectures in remote neurological monitoring. Code and anonymized structured data are publicly available: https://github.com/saiful1105020/park_video_benchmarking

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1056_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: https://papers.miccai.org/miccai-2026/supp/1056_supp.zip

Link to the Code Repository

https://github.com/saiful1105020/park_video_benchmarking

Link to the Dataset(s)

https://rochester.app.box.com/s/utw6cxrodcixenks1dxfbo9kagal7djp The dataset is described in the GitHub repository. We are actively working to add more types of anonymized data and will update the repository accordingly.

BibTex

@InProceedings{IslMd_Benchmarking_MICCAI2026,
        author = { Islam, Md Saiful AND Hossain, Ekram AND Abdelkader, Abdelrahman AND Adnan, Tariq AND Mashrur, Fazla Rabbi AND Park, Sooyong AND Kumar, Praveen AND Sudais, Qasim AND Chunga, Natalia AND Shah, Nami AND Kanan, Christopher AND Freyberg, Jan AND Schneider, Ruth B. AND Hoque, Ehsan},
        title = { { Benchmarking Video Foundation Models for Remote Parkinson’s Disease Screening } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16895},
        month = {September},
        page = {pending}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents a benchmarking study of video foundation models for Parkinson’s disease screening. It evaluates multiple pretrained video models using a fixed setup with frozen backbones and a linear classifier. The goal is to compare how different architectures perform across tasks.

  • 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 dataset is covers multiple clinically relevant tasks.
    • The task-wise analysis is useful and shows how different models behave across different types of clinical tasks.
    • The work is relevant for remote screening and has a lot of practical potential.
  • 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 results are biased towards the majority class (high specificity, low sensitivity). The paper only briefly considers oversampling and does not explore other balancing strategies (e. g. , undersampling), so it is unclear whether this is a modeling issue or a data issue. 2.The recording protocol is not very clearly described. While they mention a web-based setup and instructions, it is not fully clear how standardized the recordings are (camera position, distance, compliance with the task), especially in a remote setting. 3.The setup is quite constrained (frozen models + linear head). This is fine for benchmarking, but it also limits the conclusions. It is unclear how these models would perform with proper adaptation. 4.The choice of sustained vowels is not entirely clear. Typically, phonation tasks use vowels such as /a/, /i/, /u/ to cover the vowel space. Here, the paper uses /a/, /e/, /o/, but does not explain why these were selected. Given that vowel choice relates to articulator configuration and coverage of the vowel space, some justification would be useful. 5.It is also unclear how informative the phonation tasks are in this setting. Sustained phonation mainly reflects vocal fold behavior, which is not directly visible in video. While some articulatory movement (e. g. , lips) can be captured, this is not discussed in detail. The paper does not clearly explain why these tasks are expected to be useful in a video-only setup, or what their limitations are. More generally, since this is a benchmarking paper, the analysis should go deeper into why certain tasks work or do not work. At the moment, the paper mainly reports results without sufficiently discussing the underlying reasons or limitations.

  • 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    The paper presents a solid benchmarking study with a large dataset and a clear experimental setup. The evaluation across models and tasks is systematic. However, the results are biased towards the majority class (high specificity, low sensitivity), which is problematic for screening and not sufficiently addressed. The recording protocol is not fully clear, and it is unclear how standardized the data is. In addition, the analysis is somewhat limited. For a benchmarking paper, there is little discussion of why certain tasks (e.g., phonation) are informative or what their limitations are.

  • 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 #2

  • Please describe the contribution of the paper

    The paper’s main contribution is a large-scale, clinically structured benchmark of video foundation models (VFMs) for remote Parkinson’s disease screening, including a systematic evaluation of multiple state-of-the-art architectures across diverse clinical tasks, and an analysis showing that model performance is strongly task- and domain-dependent, providing practical guidance for selecting appropriate models in different screening scenarios.

  • 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.

    2.Strong empirical scale and rigor: The study evaluates 7 state-of-the-art VFMs on a large dataset (1,888 participants, 32k+ videos) across 16 clinically meaningful tasks. 2.Clinical relevance: The work addresses an important problem—remote Parkinson’s disease screening—with clear real-world applicability. 3.Insightful and interpretable analysis: The paper provides a valuable task-wise and domain-wise comparison, showing that performance is strongly dependent on the model architecture and clinical task, offering practical guidance for model selection.

  • 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.

    Limited methodological novelty: The paper is primarily a benchmarking study and does not introduce a novel model or algorithmic contribution. While the evaluation is thorough and well-executed, the technical contribution is incremental, as it relies on existing video foundation models with a simple downstream classifier rather than proposing new learning strategies tailored to the medical domain. Label quality and dataset limitations: The dataset includes a significant portion of self-reported PD diagnoses, which may introduce label noise and affect the reliability of the evaluation. Although the authors acknowledge this limitation, the paper does not fully analyze how this variability impacts performance or whether results differ between clinically confirmed and self-reported cases. Limited clinical readiness due to moderate sensitivity: While the reported AUC and specificity are relatively strong, sensitivity remains moderate (around 43–57%), which limits the effectiveness of the approach as a screening tool. This gap suggests that the current setup may not yet be sufficient for reliable clinical deployment and that further calibration or integration of multiple modalities may be necessary.

  • 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    The paper presents a well-executed and large-scale benchmark for remote Parkinson’s disease screening using video foundation models, with strong empirical rigor and clinically meaningful task-wise analysis. Its main strengths lie in the scale of the dataset, the systematic evaluation across multiple architectures and tasks, and the practical insights into task-dependent model performance. However, the contribution is primarily empirical, with limited methodological novelty, and is affected by dataset-related limitations such as partially self-reported labels. Additionally, the relatively low sensitivity constrains its immediate clinical applicability.

  • 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.

    The paper is still primarily an empirical benchmarking study rather than a methodological contribution, since it evaluates existing video foundation models with frozen embeddings and a simple downstream classifier. The use of self-reported PD labels may still introduce label noise, although the rebuttal partially addresses this by stating that no significant difference was observed between clinically confirmed and self-reported cases. In addition, the reported sensitivity remains moderate, limiting the immediate clinical readiness of the approach as a screening tool.

    However, after reading the authors’ rebuttal, my final recommendation is Accept. The authors addressed several important concerns by clarifying that the data split was performed strictly at the participant level, that the upper-limb motor tasks followed MDS-UPDRS instructions, and that the protocol was co-designed with MDS-certified neurologists. They also clarified their data-sharing plan, including the release of extracted VFM embeddings and anonymized metadata, which strengthens the reproducibility value of the work.

    Overall, I consider the paper suitable for acceptance because it provides a large-scale, clinically relevant, and systematic benchmark for remote Parkinson’s disease screening. The remaining issues should be clearly discussed in the final version, but they are not sufficient to change my recommendation.



Review #3

  • Please describe the contribution of the paper

    The authors conducted a robust benchmarking of various video foundation models across a very large collection of videos comprised of both PD and non-PD subjects. These videos included 16 different tasks for 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.

    The paper evaluated 7 video foundation models across a variety of frameworks. The foundation models were used for feature extraction and then a simple neural network was training as a binary classifier. The testing was completed across 30 different iterations with different seeds.

  • 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.

    As a clinical binary classifier, the “ground truth” includes self reported PD which is a severe limitation towards the interpretation/comparison of the performance between these 7 models

    Additionally, there are 32k videos across 1888 patients. Did the data split account for unique patients? Otherwise there may have been data leakage at the patient level across these videos

    The fidelity of the clinical tasks is also not clear as the videos are described to be “inspired by the MDS-UPDRS”. There are very specific instructions when trying to apply the MDS-UPDRS for clinical assessment, so much so that the MDS organization offers certification training. Were the instructions to perform these movements true to the original task instructions/guidelines? Otherwise it is unclear what the tasks actually represent. Additionally many of the tasks are not part of the MDS-UPDRS such as facial expression and the voice tasks so the model comparisons are more difficult to interpret.

    A major missing context is the medication state of the PD videos. In the medication ON state, many times the motor impairment is significantly reduced and can be close to normal in the early stages of PD.

    It is also not clear in the non-PD group how homogeneous this population is. For example, is it possible for subjects in the non-PD group to have atypical parkinsonisms or previous neurological damage that causes parkinsonism symptoms but without the diagnosis of Parkinson’s disease?

  • 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    This is a robust framework and unique dataset to benchmark video foundation models but there are strong concerns about the fidelity of the dataset. This significantly affects the interpretation of model performance comparison

  • 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.

    The authors have adequately responded to my comments and concerns. However I would like to request that they tone down the language of the conclusion, specifically “Our results suggest that frozen embeddings from large pretrained models can capture clinically meaningful signs, even without task-specific fine-tuning.” given the low sensitivity of the models (great point made by R1)



Author Feedback

We thank the Reviewers for their thoughtful feedback and for recognizing the substantial scale and contributions of our benchmarking study. ​

1.Tasks and instructions (R3): The upper-limb motor tasks map directly to MDS-UPDRS and follow the exact instructions. Speech and expression tasks are derived from items 3.1 and 3.2, and have already been validated in the literature (e.g., [1]). The protocol and instruction set were co-designed with three MDS-certified, expert neurologists.

We automatically discarded videos with obvious quality issues (e.g., abnormal durations, critical body parts obscured) and internally validated data quality by manually inspecting ~1,000 videos.​

2.Medication state (R3): We instructed participants to record videos in the OFF state to ensure medications did not mask symptoms. However, we did not enforce compliance due to safety considerations (medication withdrawal). Therefore, occasional ON state recordings are possible and may constitute label noise. We will include this as a limitation.​

3.Data split (R3): Data partitioning was conducted strictly at the participant level. Unique IDs for all participants were randomly partitioned into Training, Validation, and Test sets (Sec 2.4). All videos belonging to a specific participant were assigned exclusively to that participant’s designated set, ensuring zero leakage.

4.Self-reported PD (R2,3): We acknowledge that self-reporting can introduce label noise. We mitigated this by statistical reasoning, internal validation, and alignment with the literature. Given the low incidence of PD (<2% among the elderly), the chance of undiagnosed PD in the non-PD group is minimal and unlikely to significantly affect the benchmark’s metrics. Internally, we compared model performance between clinically confirmed and self-reported test cases and found no significant difference. Using self-reported labels in favor of a larger dataset is recognized in prior literature [2].

5.Choice of vowels (R1): We selected /a/, /e/, and /o/ because they are stable monophthongs that allow for sustained, steady-state articulatory postures. Unlike /i/ and /u/, which often involve constricted or subtle lip movements, these vowels require significant, distinct jaw opening and lip rounding, thereby maximizing the visual signal for PD detection.

6.Data balancing (R1): We avoided undersampling to maintain the diversity of the non-PD class, aiming to provide the linear classifier with the broadest possible distribution of healthy movement patterns. We internally evaluated several balancing techniques, including SMOTE, Random Oversampling, and Weighted Loss functions. None yielded significant performance gains.

Regarding low recall, we position this paper as an evaluation of foundation models, rather than as a ready-to-use screening tool. Consistent with our findings, prior work [3] also reports moderate recall for single tasks and shows that integrating multiple tasks leads to significant improvement.​

7.Detailed analysis (R1): Please note that we have provided relevant discussion in Sec 3.1 on the alignment of our findings with clinical expectations. We further hypothesized in Sec. 3.2 that model architecture and pre-training datasets may also limit model capacity. We will expand these discussions with additional qualitative and comparative analyses in the final manuscript.

8.Data sharing: To enable researchers to reproduce our findings, we will release the extracted VFM embeddings, along with anonymized metadata (demographics and clinical labels), for all 32,000+ videos.

While we cannot share raw patient videos, we will facilitate academic groups’ access to run their own feature-extraction scripts on our secure local servers (upon data usage agreement). Note that our codebase is already publicly available.

References: 1.https://ai.nejm.org/doi/full/10.1056/AIoa2400950 2.https://www.nature.com/articles/s43856-026-01606-6 3.https://doi.org/10.1609/aaai.v39i27.35031




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.

    This work presents an important benchmarking paper in the domain of Parkinson’s disease assessment. Specifically, the authors have collected a huge dataset of ~32K videos from ~1.9K patients and have benchmarked all the latest video foundation models for the task of parkinson’s disease assessment.

    All the three reviewers have agreed to the major contribution this paper makes in the assessment of parkinsons’ disease especially for the remote assesment, as the videos are recorded through webcam. The authors have already provided some samples in the supplementary, so it would be a great addition to the community.

    However, there are certain concerns from the reviewers that need to be addressed and clarified before making a decision on this paper. Please refer to the reviewer’s comments for more details. I, in particular would like the authors to pay specific emphasis to the below comments:

    1.Fidelity of the clinical tasks (R#3): Were the instructions to perform these movements true to the original task instructions/guidelines based on the MDS-UPDRS? How did the authors validate this?

    2.Medication state context (R#3): Medication state is an important context that needs to be considered. This was missing.

    3.Data split (R#3): How did the authors ensure no data leakage amongst the patients? Please provide details

    4.Self-reported PD (R#2, R#3): Self-reported PD diagnoses may introduce label noise and affect the reliability of the evaluation. How do the authors deal with this during evaluation?

    5.Choice of sustained vowels (R#1): The authors use /a/, /e/, /o/, but does not explain why these were selected.

    6.Biased towards majority class (R#1): The paper only briefly considers oversampling and does not explore other balancing strategies. Any particular reason behind this?

    7.Detailed analysis (R#1): The analysis should go deeper into why certain tasks work or do not work. At the moment, the paper mainly reports results without sufficiently discussing the underlying reasons or limitations.

    8.Data sharing: Can the authors provide a general comment on how much of the dataset will be shared openly with the community?

  • 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.

    This is definitely an important benchmark and resource to the community working in the domain of developing AI for parkinson’s disease assessment, prognosis and tracking.

    During the rebuttal, the authors have clarified and convincingly answered the questions raised by the reviewers during their first round. The authors have demonstrated their accurate guideline grounded approach in creating the dataset, the protocol design included expert clinicians, and also clarified their open resource plans. Two of the reviewers have then unanimously voted for acceptance after the rebuttal stage.

    I concur with the reviewers with the final decision. However, it would have been great to see if the raw videos dataset release, but considering the sensitive nature of the videos, it is understandable. What I’d suggest is instead of just releasing the VFM embeddings, I think the authors should also release the 2D pose keypoints, optical flow, surface normals, segmentation masks or any other relevant information that masks the privacy but captures the nuances of the motion. It would be an incredible resource for the community at large in the future.



Meta-review #2

  • 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.

    The initial reviews recognized the relevance and empirical scale of the work, but raised several important concerns. The rebuttal addressed several of these concerns in a substantive way. The authors clarified that the data split was performed strictly at the participant level, reducing the concern about leakage. They also clarified that the upper-limb motor tasks follow MDS-UPDRS instructions, that the protocol was co-designed with MDS-certified neurologists, and that quality control included automatic filtering and manual inspection of approximately 1,000 videos. The rebuttal also clarified the intended OFF-state recording instruction, while appropriately acknowledging that compliance could not be enforced for safety reasons. The explanation of the selected vowels and the authors’ rationale for avoiding undersampling help address R1’s concerns, although the clinical informativeness of some video-only speech/phonation tasks should still be discussed with more caution. Some weaknesses remain. The contribution is primarily a benchmark rather than a new method. The frozen-backbone plus shallow classifier protocol is appropriate for benchmarking representation quality, but it does not establish the performance ceiling of adapted or task-specific models. The low sensitivity, despite reasonable AUC and high specificity, means that the current system should not be framed as clinically ready for screening. The use of self-reported PD labels, possible medication-state variability, demographic imbalance, and limited information on non-PD neurological comorbidities remain important threats to clinical interpretation. The rebuttal partially mitigates these issues but does not remove them. In weighing the reviews, I give substantial weight to the post-rebuttal updates because they are specific and tied to the main technical and clinical concerns. Overall, the paper is accepted because it provides a systematic, clinically relevant, and community-useful benchmark for video foundation models in remote neurological assessment.



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.

    The authors present a novel dataset of videos for Parkinson’s disease evaluation and conduct an exhaustive benchmark of video FMs on it. The reviewers previously raised weaknesses in dataset quality (dataset split, medication status, concordance with MDS-UPDRS instructions) and in results (potential uncorrected biases with high specificity and low sensitivity). During the rebuttal phase, the authors clarified these points and appeared to have addressed all reviewers’ comments, particularly regarding dataset quality, and they clearly positioned their paper as a benchmark/application study rather than a methodological contribution. Therefore, I recommend acceptance of this paper to the MICCAI main conference.



back to top