A Vision-Based Algorithm for Assessing Head and Hand Tremor: Development and Validation Against IMU Sensors

Round 1

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

This study developed an algorithm for contactless tremor detection from 2D video recordings by analyzing the center of mass motion, implemented in open-source software TremAn3. It extracted motion data from hand and head videos, quantified tremor via spectral analysis (peak power and frequency), and validated against IMU-based accelerometry in 30 participants. The method showed moderate to good agreement with IMUs, offering a viable alternative for telemedicine and research, eliminating the need for direct sensor attachment.

The current version has issues including:

1. Part of the content in Highlights needs to be deleted, such as "The method supports simultaneous analysis of multiple body regions" which is a feature that many similar methods have; 'Video based therapy data were compared with IMU based measurements' only means a comparison was made, and no innovation or features were seen.
2. Suggest further quantifying the description of experimental results in the abstract and providing the accuracy of the evaluation or other quantitative results.
3. Suggest further explanation of the data used in the experiment, including data volume and format.
4. The experimental part lacks comparison with existing methods, making it difficult to evaluate the relative effectiveness of the proposed method in this paper.

Author Response

This study developed an algorithm for contactless tremor detection from 2D video recordings by analyzing the center of mass motion, implemented in open-source software TremAn3. It extracted motion data from hand and head videos, quantified tremor via spectral analysis (peak power and frequency), and validated against IMU-based accelerometry in 30 participants. The method showed moderate to good agreement with IMUs, offering a viable alternative for telemedicine and research, eliminating the need for direct sensor attachment.

Response: Thank you for this clear and accurate summary of our work. We appreciate your positive assessment of the proposed contactless approach and its validation against IMU-based measurements, as well as your recognition of its potential value for telemedicine and research without the need for direct sensor attachment. All revisions made in response to your  feedback are marked in cyan in the revised version for easy reference.

The current version has issues including:

1. Part of the content in Highlights needs to be deleted, such as "The method supports simultaneous analysis of multiple body regions" which is a feature that many similar methods have; 'Video based therapy data were compared with IMU based measurements' only means a comparison was made, and no innovation or features were seen.
   Response:  We have removed suggested highlights.
2. Suggest further quantifying the description of experimental results in the abstract and providing the accuracy of the evaluation or other quantitative results.
   Response: We have revised Abstract to include the key evaluation metrics.
3. Suggest further explanation of the data used in the experiment, including data volume and format.
   Response: We have added description of data, including its volume and format, into the “ Materials and Methods” section.
4. The experimental part lacks comparison with existing methods, making it difficult to evaluate the relative effectiveness of the proposed method in this paper.
   Response: We agreed that a comparison with existing video-based tremor quantification methods could further strengthen the assessment of relative effectiveness. However, our primary objective was to robustly validate the proposed video-based algorithm against IMU-derived tremor metrics, which represent the current reference standard for quantitative tremor assessment. Given the heterogeneity and implementation-specific nature of alternative video/markerless approaches (e.g., optical flow, Eulerian magnification, pose estimation), a direct head-to-head comparison was beyond the scope of this technical proof-of-concept study and was identified as future work.

Reviewer 2 Report (New Reviewer)

Comments and Suggestions for Authors

The work is well conducted, with an original contribution and a clear presentation. The methods are appropriate for the stated objectives, and the limitations are largely acknowledged and discussed. However, some technical clarifications, improvements in the presentation of results, and further discussion of methodological limitations are necessary before publication. Below is a point-by-point list of my observations. I hope these comments will be useful to the authors.

Lines 30–32: Please consider adding some numerical values to make the abstract quantitatively clearer and more informative.

Lines 124–126: The lack of synchronization between video and IMU signals is correctly mentioned, but it deserves more emphasis at this stage as a potential limitation for direct comparison. It would be helpful to briefly note here that this prevents the analysis of temporal or phase consistency between modalities.

Lines 149–159: The manual selection of ROIs is clearly described. However, it would be beneficial to specify how much time this process typically required and whether intra- or inter-operator repeatability was evaluated. If such evaluation was not performed, a brief indication that repeatability assessment is planned for future studies, as mentioned in the Discussion, could strengthen the justification.

Lines 193–194: The exclusion of frequencies below 2 Hz and above 12 Hz is understandable to avoid artifacts and noise, but this choice might also exclude certain atypical tremor patterns. It would be advisable to better justify this filtering decision based on the characteristics of the studied population.

Lines 226–232: The PPF measurement for the head shows a very low ICC value. While this issue is discussed later in the manuscript, it might be appropriate to anticipate this point here for completeness. It could be useful to state at this point that the PPF for head tremor appears unreliable, possibly due to the specific biomechanical features of cephalic tremor.

Lines 267–270: The limitation of the analyzed frequency band should be explicitly related to the generalizability of the findings. In particular, this may limit applicability to tremors with lower frequencies (e.g., Parkinsonian tremor). This could be acknowledged as an additional limitation.

Lines 325–336: Same comment as above: while the manual selection of ROIs is justified, no systematic evaluation of intra- or inter-operator variability was performed. As a suggestion for future work, the authors might consider proposing a reliability study involving multiple raters, or the integration of an automatic tracking module.

Author Response

The work is well conducted, with an original contribution and a clear presentation. The methods are appropriate for the stated objectives, and the limitations are largely acknowledged and discussed. However, some technical clarifications, improvements in the presentation of results, and further discussion of methodological limitations are necessary before publication. Below is a point-by-point list of my observations. I hope these comments will be useful to the authors.

Response: Thank you for your positive and constructive assessment. We appreciate your recognition of the originality, clarity, and methodological appropriateness of our work. We also value your detailed point-by-point comments, which we have addressed through technical clarifications, improved presentation of results, and an expanded discussion of methodological limitations. All revisions made in response to your feedback are marked in magenta in the revised version for easy reference.

Lines 30–32: Please consider adding some numerical values to make the abstract quantitatively clearer and more informative.

Response: Thank you for the suggestion. We have revised Lines 30–32 of the abstract to include key quantitative results. This change is also consistent with Reviewer 1’s comment.

 

Lines 124–126: The lack of synchronization between video and IMU signals is correctly mentioned, but it deserves more emphasis at this stage as a potential limitation for direct comparison. It would be helpful to briefly note here that this prevents the analysis of temporal or phase consistency between modalities.

Response: Thank you for highlighting this point. We have revised the manuscript to explicitly state that the lack of synchronisation prevented any analysis of temporal alignment or phase consistency between modalities.

 

Lines 149–159: The manual selection of ROIs is clearly described. However, it would be beneficial to specify how much time this process typically required and whether intra- or inter-operator repeatability was evaluated. If such evaluation was not performed, a brief indication that repeatability assessment is planned for future studies, as mentioned in the Discussion, could strengthen the justification.

Response: We thank the reviewer for this suggestion. We have clarified that intra- and inter-operator repeatability was not evaluated in the present study and stated that a dedicated repeatability assessment is planned in a future study.

 

Lines 193–194: The exclusion of frequencies below 2 Hz and above 12 Hz is understandable to avoid artifacts and noise, but this choice might also exclude certain atypical tremor patterns. It would be advisable to better justify this filtering decision based on the characteristics of the studied population.

Response: Thank you for this comment. We have clarified in the revised manuscript that our primary aim was to validate the video-based algorithm (not to comprehensively analyse atypical tremor patterns). Therefore, the evaluated frequency range was aligned with the expected tremor frequencies in our ET/dystonic cohort, and we noted that atypical components outside this range were not systematically assessed.

 

Lines 226–232: The PPF measurement for the head shows a very low ICC value. While this issue is discussed later in the manuscript, it might be appropriate to anticipate this point here for completeness. It could be useful to state at this point that the PPF for head tremor appears unreliable, possibly due to the specific biomechanical features of cephalic tremor.

Response: We have revised the manuscript to briefly flag at this stage that head PPF showed poor agreement (very low ICC) and therefore appeared unreliable in our data, likely reflecting the biomechanical characteristics of cephalic tremor and the limitations of a 2D video-based CoM approach for head motion. A more detailed explanation was retained in the Discussion section.

 

Lines 267–270: The limitation of the analyzed frequency band should be explicitly related to the generalizability of the findings. In particular, this may limit applicability to tremors with lower frequencies (e.g., Parkinsonian tremor). This could be acknowledged as an additional limitation.

Response: We have updated the manuscript to explicitly link the chosen analysis band to the generalizability of our findings.

Lines 325–336: Same comment as above: while the manual selection of ROIs is justified, no systematic evaluation of intra- or inter-operator variability was performed. As a suggestion for future work, the authors might consider proposing a reliability study involving multiple raters, or the integration of an automatic tracking module.

Response: Following your comment above on manual RoI selection, we revised the Discussion section to state that intra- and inter-operator repeatability was not evaluated in this study and that a dedicated multi-rater repeatability study is planned.

Reviewer 3 Report (New Reviewer)

Comments and Suggestions for Authors

This manuscript presents a clear and technically sound approach for contactless tremor quantification using standard 2D video recordings. The proposed center-of-mass–based algorithm is simple, interpretable, and well validated against IMU-derived tremor metrics in a clinically relevant cohort. The study is timely and relevant, particularly in the context of telemedicine and low-cost remote assessment of movement disorders.

The main strengths of the work are the methodological transparency, the direct comparison with established sensor-based measurements, and the open-source availability of the algorithm. The discussion appropriately acknowledges key limitations, including challenges related to head tremor characterization, manual ROI selection, and recording conditions.

A brief clarification of the proof-of-concept nature of the sample size and minor language polishing in the Discussion would also improve clarity. Overall, this is a solid contribution and is suitable for publication after minor revisions.

Author Response

This manuscript presents a clear and technically sound approach for contactless tremor quantification using standard 2D video recordings. The proposed center-of-mass–based algorithm is simple, interpretable, and well validated against IMU-derived tremor metrics in a clinically relevant cohort. The study is timely and relevant, particularly in the context of telemedicine and low-cost remote assessment of movement disorders.
The main strengths of the work are the methodological transparency, the direct comparison with established sensor-based measurements, and the open-source availability of the algorithm. The discussion appropriately acknowledges key limitations, including challenges related to head tremor characterization, manual ROI selection, and recording conditions.
A brief clarification of the proof-of-concept nature of the sample size and minor language polishing in the Discussion would also improve clarity. Overall, this is a solid contribution and is suitable for publication after minor revisions.
Response: 
Thank you for your positive and constructive assessment. We appreciate your recognition of the clarity, technical soundness, and interpretability of our center-of-mass–based approach, as well as its validation against IMU-derived metrics and relevance for telemedicine and low-cost remote assessment. 
We also thank the reviewer for the helpful suggestions to improve clarity. We have clarified in the revised manuscript that the sample size reflected a proof-of-concept validation study (marked in green in the revised version for easy reference). We have performed minor language polishing in the Discussion as you suggested.

Round 2

Reviewer 1 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

The authors have already provided a good answer to my concerns. I don't have any other questions.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors present an interesting approach to assessing head and hand tremors through a vision-based algorithm, offering a promising alternative to traditional methods. By utilizing video recordings to detect tremors based on the motion of the center of mass, the authors demonstrate significant potential for advancing diagnostic and monitoring techniques in movement disorders.

However, there are areas that warrant further exploration to enhance the study's impact.

1. Methodology Clarity: Further details on the video recording setup and the specific calculations for peak tremor power and frequency would enhance replicability.
2. Validation Strengths: The validation against IMU data is robust, but comparing with other existing methods could strengthen the study's impact. Additionally, testing across diverse patient groups and conditions would demonstrate broader applicability.
3. Clinical Application Potential: The algorithm shows promise for telemedicine, particularly in remote settings. Future work should focus on real-time processing capabilities and integration into clinical diagnostic tools to enhance practicality and accessibility.

Reviewer 2 Report

Comments and Suggestions for Authors

Please see PDF

Comments for author File: Comments.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

 The manuscript presents a contactless, vision-based approach for the estimation of head and hand tremor. The authors utilize Inertial Measurement Units (IMUs) as the gold standard to validate their approach, which is primarily based on Computer Vision techniques. Tremor is estimated by tracking the movements of a Region of Interest (ROI), which is manually defined in the initial frame of the recorded video. This area defined by the ROI is subsequently processed across video frames to extract frequency and power information to be compared against the equivalent metrics derived from the accelerations of IMU data.

The topic is clinically relevant, as tremor affects various neurological pathologies impairing activities of daily living (ADLs). The main advantage of a contactless, vision-based approach is the quantification of tremor features without interfering with the movement itself, as could potentially occur with wearable sensors applied to the body. However, the setup utilized for validation, which includes 3 wearables, may inherently limit this key advantage. It would have been more methodologically appropriate to use a motion capture (MoCap) system as the gold standard, thereby enabling a comparison between two distinct contactless vision-based approaches.

In addition to this general observation, the manuscript lacks clarity in several key sections and important details. Furthermore, the methodological approach appears simplistic and exhibits poor robustness, as it is based on the manual definition of a Region of Interest (ROI) in the initial frame, which is then maintained throughout the entire task. This is particularly relevant when compared to existing studies based on more sophisticated hand tracking models. Specific comments follow:

• The Introduction (and the References section) should be expanded to include recent studies that apply hand tracking methods to tremor assessment.
• The authors should specify the innovative contributions of their proposed approach in greater detail. They should also clearly articulate the advantages of using such a simplistic methodology compared to current state-of-the-art methods.
• Lines 100-101: How was the synchronization between the 3 IMUs and the video stream handled? The authors should provide further details on this critical aspect of the methodology.
• Lines 103-104: Did the participants assume the SWing pose before the signal recording began, or after the recording was already initiated? The authors should clarify these protocol steps, as initiating the pose after the recording starts could introduce significant motion artifacts.
• Line 114: The authors should clarify whether the manually drawn ROI on the first frame encompasses both hands, or if separate ROIs were defined for the right and left hands. Moreover, the authors state that the ROI is defined in the first frame and remains fixed for the entire recording. This seems a significant limitation. Could this constraint be too stringent, especially for subjects with tremor? Is there not a risk that the hands move outside of this pre-defined ROI during the 20-22 second task? What was the size (e.g., in pixels) of the ROI? Was this ROI size standardized across all analyzed videos, or did it vary between recordings? What was the subject's distance from the camera? Was this distance kept consistent for all participants? The authors must provide more comprehensive details regarding the standardization of the video acquisition procedure, as these factors could significantly influence the algorithm's performance and the results.
• The video processing methodology for head tremor estimation is entirely missing from the manuscript. The authors do not provide any details regarding this aspect.
• Figure 2: Considering the setup shown in Figure 1, the hand is viewed laterally (in profile), not frontally. Figure 2 should show an example of the actual ROI processing for this specific view. Furthermore, based on its design, the ROI (Figure 2) might include portions of the face. Since the face could be affected by head tremor, this inclusion could introduce significant confounding signals (noise) into the ROI. The authors must clarify how their method isolates only the hand movement within the ROI for the calculation of the Center of Mass (CoM).
• Lines 144-145: The authors state that three PSDs are estimated for the acceleration signals (one for each IMU) and two for the CoMs (from ROIs). This seems a discrepancy, as it is unclear why only two PSDs are computed for the video analysis. Logically, there should be one ROI for the head and one for each hand (i.e., three ROIs), unless a single ROI was used for both hands. This point reinforces the significant lack of clarity in the methodological description, as it remains ambiguous how the ROIs for the specific areas of interest (head vs. hands) are drawn.

• Line 153: The authors report problems in six acquisition sessions due to artificial light (flickering) within the experimental environment. Consequently, the authors must clarify the specific environmental conditions (particularly lighting) under which the other trials were performed. Lighting conditions are a well-known and critical challenge for Computer Vision algorithms. What precautions did the authors take to ensure the robustness of their methodological approach against such lighting variations? Furthermore, what is the potential impact of conducting trials in an uncontrolled environment with heterogeneous lighting conditions on the overall performance and, consequently, on the results?
• Lines 154-155: The authors apply cut-off thresholds (2-4 Hz and 10-12 Hz) to remove anomalous peaks attributed to artificial lighting. This choice seems problematic, as 4 Hz is a frequency compatible with pathological tremors. The authors must provide a justification for selecting this 4 Hz boundary, as this could inadvertently filtering out relevant frequency components of the tremor signal itself.
• Table 1: In addition to the ICC and MAE values, the authors should provide, for completeness, the descriptive statistics (i.e., mean and standard deviation) for the peak power parameter for both IMUs and ROIs across all three regions of interest. Furthermore, the unit of measurement for the MAE must be specified. Perhaps preferably, the MAE should be reported as a percentage value (e.g., relative MAE): it is impossible to determine whether an MAE value > 8 (for example) represents a large or a small error. Including the mean and standard deviation of the peak power values is essential to properly interpret and contextualize the magnitude of this absolute error.
• Table 2: As for Table 1, the authors should provide the descriptive statistics (mean and standard deviation) for the peak frequency parameter for both the IMU and video-derived measurements for all three regions of interest. The unit of measurement for the MAE must also be specified (presumably Hz). Most importantly, the authors must provide a plausible explanation for the large error reported for the head (MAE > 2 Hz), which contrasts with the lower error reported for the hands. Given that the methodological approach for video data analysis is identical for all regions, what is the source of this significant discrepancy?
• Figure 3: The ranges of the X and Y axes across the various plots for Peak Frequency (PF) and Peak Power (PP) should be aligned for easier comparison. Furthermore, some data points appear to be off-scale (e.g., in plots A, C, and