From RGB-D to RGB-Only: Reliability and Clinical Relevance of Markerless Skeletal Tracking for Postural Assessment in Parkinson’s Disease

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This is a timely and clinically relevant validation study comparing an RGB-only markerless pipeline (Google MediaPipe Pose, multiple complexity levels) with an RGB-D reference (Azure Kinect) in 40 people with Parkinson’s disease during a 60-s quiet standing task captured with a synchronized frontal+lateral setup.  My overall recommendation would be major revision, mainly to improve methodological rigor in validation metrics, strengthen statistical control, and make the work more reproducible and tele-rehab ready. 

Major comments: 

1. The manuscript relies primarily on Pearson correlation (and Wilcoxon tests) to support “agreement/consistency” between MediaPipe variants and the KIN_3D reference. However, correlation does not measure agreement and may remain high even in the presence of systematic bias, which is also acknowledged for several parameters.
   Please complement correlation with additional agreement metrics for the key angles deemed clinically usable, such as RMSE/MAE, Bland–Altman analysis (bias and limits of agreement), and/or ICC (clearly specifying the ICC model and assumptions).

2. The Discussion refers to the RGB-D system as “ground truth” based on previous validation against optical motion capture. While this supports its use as a reference, the term “ground truth” may be too strong for a clinical validation study without a gold-standard system in the present dataset.
   Consider using “reference system” throughout (unless a gold-standard subset is added), and briefly discuss the expected uncertainty/limitations of KIN_3D in static posture assessment, particularly in scenarios with potential occlusions or the presence of operators in the field of view.

3. The protocol includes both eyes-open and eyes-closed quiet-standing tasks. It is currently unclear whether analyses were performed on one condition only, averaged across conditions, pooled, or analyzed separately. So clearly state the handling of EO/EC in each analysis and, if applicable, consider reporting whether agreement differs between conditions.

4. Given the potential relevance for tele-assessment and telerehabilitation, the Discussion would benefit from a concise subsection on real-world deployment.
   You could  briefly outline practical requirements for home/remote use, including minimal camera placement constraints, quality-control indicators (e.g., confidence thresholds, occlusion detection, re-identification stability), and how the proposed multi-person handling could generalize when caregivers enter the scene.  

Minor comments 

1. Lines 67–71: please add an appropriate reference to support the statement(s).

2. Lines 124–136: this content may fit better in the Discussion, or alternatively it could be moved earlier (e.g., at the end of the Introduction) to better justify the study rationale and clarify how this work differs from prior literature before stating the objectives.

3. Please add a table summarizing the main demographic and clinical characteristics of the participants. For example: age, sex, medication regimen (if available), and key clinical outcomes, reported as mean ± SD (or median [IQR] where appropriate). Currently, these data are dispersed in the text and are difficult to track.

 

 

 

Author Response

Please seethe attached PDF file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript investigates the reliability and clinical relevance of markerless skeletal tracking for postural assessment in patients with Parkinson’s disease (PD), comparing an RGB-only framework (Google MediaPipe Pose, MP) with an RGB-D reference system (Microsoft Azure Kinect, MAK). Using a synchronized dual-camera setup (frontal and lateral views), the authors analyzed 3D angular postural parameters during a 60-second static standing task in 40 PD patients. Three MP model complexities were evaluated and benchmarked against MAK across horizontal, vertical, sagittal, and joint-level angles. The study demonstrates that lower-complexity MP models (especially MP_3D_0) show strong agreement with MAK for axial and proximal postural measures, whereas the high-complexity model introduces substantial reconstruction errors. Clinically, several static postural metrics correlated meaningfully with established motor, balance, and complication-related clinical scales, and both RGB-only and RGB-D approaches successfully discriminated between postural severity clusters. Overall, the study supports the use of RGB-only markerless systems as accessible and clinically relevant tools for objective postural assessment in PD, while clearly defining their technical boundaries.

 

Major Comments and Questions for the Authors

1. Although MAK is treated as the reference system, it is not a gold-standard motion capture system. While prior validation studies are cited, it would strengthen the manuscript to more explicitly discuss how residual errors in MAK might propagate into the reported correlations and biases, particularly for distal joints.
2. The manuscript appropriately emphasizes correlation over absolute agreement; however, in a clinical context, systematic angular offsets (e.g., ~10° in sagittal trunk angles) may still be relevant. Please clarify how such offsets would affect longitudinal monitoring or inter-clinic comparisons, and whether simple calibration or normalization strategies could mitigate this issue.
3. Pearson correlation is used extensively to assess agreement between systems. Have the authors considered complementary agreement metrics (e.g., ICCs or Bland–Altman analysis) to better characterize absolute agreement and bias, especially given the clinical application?
4. The use of the Nose landmark as a proxy for head position in MP is acknowledged as a limitation. Could the authors comment on whether alternative composite head landmarks (e.g., averaging facial points) were considered, and how sensitive the results might be to this choice?
5. The observed association between more upright static posture and worse freezing of gait or ADL performance is interesting and well discussed. However, the manuscript would benefit from a clearer distinction between hypothesis-generating findings and clinically actionable conclusions, particularly to avoid misinterpretation by readers less familiar with PD phenotypes.
6. The study was conducted in controlled laboratory settings with careful camera placement and synchronization. Please expand the discussion on how robust the proposed MP-based approach would be under less controlled home or telemedicine conditions, and which parameters are most likely to remain reliable.
7. Sample Size and Severity Distribution
   While the cohort size is reasonable for a technical–clinical validation study, the imbalance across postural severity levels may affect some correlations. Please clarify how many participants fell into each severity cluster and whether this imbalance could influence PCA-based separation.

Minor Comments and Suggestions

• The manuscript is very dense in terms of reported parameters. A concise summary table highlighting clinically recommended angles for RGB-only use would improve usability for clinicians and applied researchers.
• Some tables (e.g., Tables 5–8) are information-rich but challenging to interpret quickly. Consider visually highlighting parameters that meet predefined reliability thresholds.
• Please ensure consistent use of terminology when referring to sagittal angles derived from frontal vs. lateral views to avoid confusion.

 

Quality of the English

The overall quality of the English is high, with clear structure, precise terminology, and appropriate academic tone throughout. The manuscript is generally well written and easy to follow despite its technical complexity. Minor issues include occasional long or overly dense sentences and small grammatical inconsistencies, but these do not impede comprehension. The discussion section is particularly strong in articulating nuanced clinical interpretations. Only light copyediting is recommended, mainly to improve conciseness and readability rather than correctness.

 

Author Response

Please see the attached PDF file.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1. Summary and Key Contributions

The manuscript entitled “From RGB-D to RGB-only: Reliability and Clinical Relevance of Markerless Skeletal Tracking for Postural Assessment in Parkinson’s Disease” presents a comprehensive technical and clinical validation study comparing 3D human pose estimation (HPE) frameworks for postural analysis in Parkinson’s Disease (PD) patients.

Key contributions of the work are:

1. Technical validation: Direct comparison of three Google MediaPipe Pose (MP) complexity levels (0, 1, 2) against a Microsoft Azure Kinect (MAK) reference system using a synchronized dual‑camera setup, enabling precise evaluation of RGB‑only versus RGB‑D tracking.
2. Multi‑view advantage: Demonstration that adding a lateral (Sub) camera view significantly improves sagittal‑angle estimation with MP models, overcoming inherent depth‑estimation limitations of single‑view RGB‑only frameworks.
3. Clinical relevance: Correlation of objective postural parameters with standardized clinical scales (MDS‑UPDRS, BBS, NFOG‑Q, etc.) and successful discrimination of postural‑severity clusters, establishing MP as a viable tool for digital assessment of axial postural abnormalities in PD.
4. Practical recommendation: Identification of lower‑complexity MP models (MP_3D_0, MP_3D_1) as the most reliable for clinical use, while warning against the high‑complexity model (MP_3D_2) due to skeletal‑reconstruction errors.
5. Methodological innovation: A robust, synchronized dual‑camera protocol that allows fair comparison between depth‑based and vision‑only pose estimation, providing a blueprint for future validation studies.

Overall, the study makes a strong case for transitioning from specialized RGB‑D hardware to more accessible, software‑based MP solutions for objective postural assessment in clinical and potentially remote settings.

 

1. Detailed Evaluation of Methodology, Analyses, and Conclusions

Strengths

1. Study design and synchronization: The use of a daisy‑chained, dual‑MAK setup ensures temporal alignment of frontal and lateral views, a critical requirement for comparing MP (RGB‑only) and MAK (RGB‑D) outputs.
2. Comprehensive parameter set: Horizontal, vertical, sagittal, and 3D joint angles are computed, along with segment lengths and symmetry indices, offering a multi‑faceted postural profile.
3. Robust statistical approach:

• Use of non‑parametric Wilcoxon Signed‑Rank Test for paired comparisons.
• Pearson correlation for inter‑model and clinical‑scale associations.
• Principal Component Analysis (PCA) to visualize cluster separation based on clinical severity.

1. Clear technical‑clinical linkage: The work moves beyond pure technical validation by linking angular measurements to established clinical scales and demonstrating discrimination between posture‑severity clusters.
2. Transparent limitations: The authors explicitly acknowledge challenges in distal‑limb tracking, environmental constraints, and the need for larger cohorts and dynamic tasks.

Potential Concerns and Areas for Improvement

1. Methodological clarifications needed

• Skeletal‑model correspondence: The manuscript states that “common points (joints and landmarks)” were identified between MP and MAK, but the exact mapping (e.g., MP nose vs. MAK head joint) should be more explicitly described, perhaps in a supplementary table.
• Masking and patient‑tracking procedure: While the masking method for MP is mentioned, details on how the masking region was defined (manual vs. automated) and whether it could introduce bias in initial pose estimation are lacking.
• Resampling and filtering parameters: Resampling to 50 Hz and the choice of a 5 Hz low‑pass cut‑off are reasonable but could be further justified (e.g., citing literature on postural‑sway frequency content in PD).

2. Statistical considerations

• Multiple‑comparison adjustments: With numerous angular parameters and clinical scales tested, the risk of Type I error increases. Applying a correction (e.g., False Discovery Rate) to the correlation matrices would strengthen the validity of the reported significant associations.
• Correlation vs. agreement: Pearson correlation assesses linear relationship, not absolute agreement. Supplementing with Bland‑Altman plots for key angles (e.g., ZM_TRUNK, ZS_TRUNK) would provide a clearer picture of measurement bias and limits of agreement between MP and MAK.

3. Clinical interpretation and generalizability

• Cohort characteristics: The cohort consists of moderate‑to‑advanced PD patients (H&Y 2.5‑4). The findings may not generalize to early‑stage patients or other populations with postural abnormalities. This limitation should be emphasized in the Discussion.

4. Presentation and clarity

• Figure and table references: Some references to supplementary figures (e.g., Figures S1–S4) appear in the main text, but the supplementary material is not provided for review. Ensure all cited supplementary data is included in the submission.
• Abbreviations: While most abbreviations are defined, some less common ones (e.g., LEDD, NFOV) could benefit from a glossary or expanded explanation at first use.
• Conclusion refinement: The conclusion could more explicitly state that MP_3D_0 is recommended as the optimal trade‑off between accuracy and computational cost for clinical postural assessment in PD, based on the presented data.

 

1. Constructive Feedback and Suggestions for Authors

Major Revisions Recommended

1. Enhance methodological transparency:

• Provide a detailed mapping table between MP landmarks and MAK joints used for angle calculations.
• Describe the masking procedure (including software/tools used) and its potential impact on initial pose estimation.

1. Strengthen statistical reporting:

• Apply a multiple‑comparison correction to the correlation analyses and report adjusted p‑values or q‑values.
• Include Bland‑Altman plots for key angles comparing MP models to MAK.

1. Refine clinical interpretations:

• Explicitly state that findings are specific to moderate‑advanced PD and may not apply to early‑stage patients.
• Expand the discussion on the dissociation between static alignment and dynamic function, considering alternative mechanistic explanations.
• Briefly review MP’s performance in dynamic tasks (e.g., gait) from literature to contextualize future applications.

1. Improve manuscript clarity:

• Ensure all supplementary figures/tables referenced in the main text are provided.
• Consider adding a schematic of the angle‑calculation workflow (frontal vs. lateral) to aid reader comprehension.
• In the Abstract, explicitly mention that MP_3D_0 and MP_3D_1 are the recommended models, whereas MP_3D_2 introduces significant errors.

Minor Revisions

• Check consistency in reporting decimal places (e.g., means and standard deviations).
• Unify terminology: e.g., “MAK” vs. “KIN_3D” appears interchangeably—standardize to one preferred term.
• Proofread for minor grammatical/typo errors (e.g., “meas- urements” line 40, “disabiliy” line 741).

Author Response

Please see the attached PDF file.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed my comments satisfactorily. I have no further comments.

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for addressing the comments.