Using Wearable Accelerometers to Develop a Vertical Ground Reaction Force Prediction Model during Running: A Sensitivity Study

Round 1

Reviewer 1 Report

The paper is well written and the subject of study very interesting. I have only a few of comment that could help to improve its quality.

In Results section, maybe it would not be necessary to show a table with the data of all the runner's speed one subject. You do not use these data after that so the mean value of speed would be enough. 

The example of one of the model for the different participants from line 279 to 289 could be removed from this results section. In Fig. 4 why do not show the worst and best case as in the previous figures?

In the discussion section the reference in lines 342-343 talks about pathology. Maybe a reference with healthy subjects data will be more suitable.

In the statement of the line 370 you recommended a combination of cut-off frequencies but you have not tested them, is not it? why did not test if you saw it?

In test phase section you talk about the velocity in the accuracy of the prediction model. could you compare the faster and the slower runners best model in order to detect this speed dependence?

Would it help that the first phase of the measurements would be made at different running speeds? 

In the conclussion it should be added that the model is valid for a certain speed. Higher o lower velocities would not be well modelled. 

 

 

 

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript proposes VGRF prediction models for running using dual accelerometers and a small dataset, showcasing strong correlations and minimal errors. The authors also discuss the impact of both the number of stances and the filtering frequency on the model. Although validation is needed for broader applicability, this approach holds promise for biomechanical monitoring during running with minimal sensors.

 

In general, the experimental design and statistical analysis in this study are praiseworthy. Nevertheless, there are apprehensions pertaining to the figures, data utilization, and discussion. Consequently, the Reviewer asserts that the current version of the paper cannot be accepted, and a major revision is necessary.

 

(1)    The two peaks, passive peak and active peak, are currently illustrated only in the result, and the loading rate is merely described in textual form. To further enhance the clarity and comprehensibility of the description regarding vGRF indicators, it is recommended that the authors incorporate a schematic diagram into the manuscript.

 

(2)    In Section 2.5, the authors mention the final averaging operation. Does this imply that a single set of 5000 sample points was used for training or validation, while eight sets of 5000 sample points were utilized for testing? If such an approach was employed, it could potentially lead to a suboptimal utilization of the available data. Consequently, it raises concerns about the potential impact on the effectiveness of the training process.

 

(3)    In Section 3.2, it is mentioned that only the model developed with 10 steps and with accelerations filtered under 10Hz did not include the medio-lateral tibia acceleration. It is recommended to elucidate the reason for this decision and assess whether the resultant omission is justifiable within the context of the study.

 

(4)    Given significant disparities observed in the characteristics of predicted vGRF as compared to the real measurements, including variations in numbers of peaks, it is highly recommended that the authors point out these disparities, outline the specific reasons and expound upon why the model is acceptable.

 

(5)    During the comparison depicted in Figures 3 and 4, a notable difference is observed in the ground truth data between the left and right frames. Based on the authors' statement, this difference might stem from varying cut-off frequencie. However, this situation gives rise to a significant contrast in the characteristics of the same dataset (with same stances). The reviewer doubt whether it is acceptable to label these data as "real VGRF". Consequently, referring to ground truth that has been processed in different ways becomes problematic. Such a comparison becomes inherently unfair and, moreover, loses its meaningfulness due to dissimilar ground truth.

 

 

(6)    In Section 4.2, the authors highlight that the accelerometer positioned on the tibia contributes to modeling the initiation of the vGRF curve, primarily through the passive peak and loading rate components. Conversely, sensors placed in the lumbar region, particularly along the longitudinal axis, offer essential information for modelling the active peak. This conclusion, however, seems to lack further detailed explanation in the preceding text. Is it based on the coefficients of the trained model?

Proofread is required to improve the language.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have well addressed my comments, and I, therefore, suggest acceptance of this manuscript.