Wrist-Based Fall Detection: Towards Generalization across Datasets

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper explores fall detection using acceleration and gyroscopic data collected from the wrist, leveraging four publicly available datasets. Thirteen statistical parameters serve as input features for two machine learning models, namely Artificial Neural Networks (ANN) and Support Vector Machines (SVM). The findings are presented alongside a comparison with existing methods. However, the contribution may be deemed less substantial due to the challenge of accurately gauging the improvement of the proposed method over diverse prior studies.

 

Does the proposed method employ the same cross-validation approach in each comparison between training and test sets as seen in previous studies?

 

In the comprehensive datasets, diverse data from different locations are included. Were the results in each comparison evaluated based on wrist’s data?

 

While an abbreviation list is provided at the end of the paper, it is essential to present the full name of each abbreviation at its initial occurrence in the text.

 

Tables 10-12 outline the comparison of results using both normalized and non-normalized data. While data normalization has proven effective in enhancing machine learning outcomes, these specific comparisons may be deemed unnecessary.

 

The Results section appears disorganized, with an abundance of tables that may potentially confuse readers. It is recommended to place particular emphasis on illustrating how the proposed method enhances the performance of fall detection.

 

The paper lacks discussions on key aspects, including the improvement of the proposed methods, limitations inherent in the proposed approach, distinctions between the use of ANN and SVM, as well as potential challenges associated with utilizing wrist data for fall detection. Integrating these aspects would contribute to a more comprehensive analysis.

Comments on the Quality of English Language

While the literature review is well-articulated, the subsequent sections lack organization and clarity in their descriptions. Consider refining and structuring the content in English to enhance coherence and understanding..

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The study is interesting with good results. Some key limitations of the study:

- Highlight the key contributions of your work and ellaborate how this work is novel compared to other related work.
- Should add more relevant work.
- Explain the method sections more - particularly how the imbalanced data has been adjusted.
- Reiterate your novelty and contribution again in the conclusion. 

Comments for author File: Comments.pdf

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for a very interesting paper.  Most of my comments pertain to the introduction, and I will leave the other reviewer/s with more knowledge of maching learning to comment on those sections.  It would be interesting to know if the sensors have any impact on fear of falling?  Please consider updating and expanding the lit review and linking it to any patient-based data about the impact of fall detection systems.

Comments for author File: Comments.pdf

Comments on the Quality of English Language

Some minor issues, please reconsider the use of the word 'elderly'.

Author Response

Please see the attachment

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Tables 8, 9, and 10 present comparisons with previous studies. Specifically, Table 8 indicates lower accuracy and recall but higher specificity compared to earlier research, suggesting a distinct prioritization in model construction. Additional details regarding the employed parameters for these comparisons are provided. Meanwhile, is it accurate to state that the proposed approach exhibits similarity between accuracy and specificity, while the recall differs significantly? Similar discrepancies are observed in Table 9, where the proposed approach exhibits higher accuracy compared to recall and specificity.

 

The revised paper still lacks discussions on key aspects, including possible mechanisms for the comparison with previous studies, the improvement of the proposed methods, limitations inherent in the proposed approach, distinctions between the use of ANN and SVM, as well as potential challenges associated with utilizing wrist data for fall detection. Integrating these aspects would contribute to a more comprehensive analysis. A independent discussion section is needed.

 

The authors mentioned “Using classic ML algorithms (SVM and ANN) and considering datasets UP Fall, UMA Fall, 528 and WEDA Fall, we introduce a new combined dataset that has over 1300 fall samples 529 and 28K ADL samples.” I don’t think this constitutes a combined study. Each dataset is independently trained and validated. The use of a confusion matrix for each dataset would provide more detailed insights into the classification numbers.

 

I suggest that the authors undertake substantial revisions and clearly articulate their responses to the reviewer's comments, incorporating necessary modifications.

Comments on the Quality of English Language

NA

Author Response

Please see attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Tha authors have addressed my comments in suficient manner.

Author Response

According to the reviewer's comment, we addressed all the issues in the revised version of the paper.

Many thanks to the reviewer for the comments.