AI-Driven Adaptive Segmentation of Timed Up and Go Test Phases Using a Smartphone

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Authors have presented a smartphone based AI driven method for TUG test. They have involved 27 volanteets for sample collection while performing stand-walk-turn-walk-sit activity chain. An android phone attached to the lower back is used to collect sensor data.

I have few concerns that must be clarified to bring the paper to a better quality. 

It is unclear why they have attached a smartphone to the lower back as it would destract the walking style of a person and will disturb the sitting activity. If it was a miniature IMU, the effect would be minimal. Also, it is not very clear why a location like the front pocket (thigh) was not considered as a data capturing location. Authors must justify why they used a larger device attached to the lower back, vs a smaller device and/or properly locating the device.

Further, the colors used in Figure 3 makes the figure unclear. Please change them so the numbers would be visible. 

 

Author Response

Dear Reviewer,
Thanks for the feedback; now your suggestions are incorporated in the revised manuscript. Find our responses to your comments in the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper is a study that utilizes a single smartphone attached to the low back (L3 position) to automatically detect the main phases of the Timed Up and Go (TUG) test—namely, sit-to-stand, walk, turn, walk, turn, and stand-to-sit—using AI. I believe it achieved good results that are highly suitable for future clinical application.

The TUG test is particularly useful for assessing fall risk, but it has issues regarding accuracy, variability due to factors like diverse patient conditions, subjectivity of the assessor, and standardization. Therefore, the study's achievement of accuracy in measuring performance time and other metrics using a smartphone is a very significant research finding. Furthermore, precisely measuring the start and end points of each TUG phase is clinically crucial.

However, a very regrettable aspect of this study is that it did not perform a comparison with video camera measurements to actual application and comparison of results in real patients, nor did it include a process for scoring key elements such as phase-specific performance capability. Simply noting that the TUG duration for patients is a few seconds longer than for healthy individuals is insufficient; recognizing the precise timing and completion of each phase requires additional algorithmic tuning. (There were similar algorithmic tuning processes when attempting to recognize the four key events of gait using AI algorithms.) Real-world validation is absolutely necessary for application to patients, and subsequent refinement and modifications are essential. (I hope the process for clinical application will be included in the discussion of this paper.)

Nonetheless, the paper's research methodology and the value of its results are relatively sound overall, and I hope that these findings will be utilized through clinical application in the paper's follow-up research.

Author Response

Dear Reviewer, Find the attached feedback to your suggestions.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This paper presents a smartphone-based system that uses AI to automatically analyze the Timed Up and Go test, which is widely used to assess mobility and fall risk. Compared to V1, the new version is much clearer and more rigorous. I only have a few remaining questions:

1. The decision to rely only on the L3 vertebra placement makes sense from a biomechanics perspective, but it limits real-world use. Most people keep their phones in a pocket or a bag. It’s unclear how well the system would work in those everyday placements, which could be a major barrier to adoption.
2. The machine learning models are trained using labels produced by the rule-based system. That means the models may just be learning the rule-based heuristics rather than the true ground truth from video. This could make the performance metrics look better than they actually are.
3. The use of standard statistical features is reasonable, but the paper doesn’t really explore whether all of them are necessary. Some basic feature analysis, correlations, feature selection, etc., could help identify a smaller and more efficient feature set.
4. The paper emphasizes interpretability, but the best-performing model is still a black box. Global feature importance doesn’t explain individual predictions, and that may limit how clinicians perceive or trust the system.
5. The system is described as practical and scalable, but there’s no analysis of runtime, memory use, or power consumption on an actual phone. For an electronics/embedded-systems audience, this is important information and feels missing.
6. The confusion matrices show consistent confusion between Walk and Turn. It would help to look more closely at when these errors occur, especially whether they happen during transition periods, so the temporal segmentation can be improved.

Author Response

Dear Reviewer, thank you for taking the time to provide a detailed review of our paper. A summary of our responses to your suggestions is provided in the attached file.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

Review electronics-3968337-peer-review-v2

The manuscript titled AI-Driven Adaptive Segmentation of Timed Up and Go Test Phases Using a Smartphone presents a smartphone-based system that automatically segments the Timed Up and Go (TUG) test into its core phases through adaptive preprocessing, sensor fusion, and supervised machine-learning models. The topic is timely and relevant, given the need for low-cost mobility assessment and the increasing evidence supporting IMU-based gait analysis. Overall, the article is clearly written, technically detailed, and structured according to standard scientific format (Introduction, Related Work, Methods, Results, Discussion, Conclusions). I have only few minor comments.

Sections introduction and Related Work are generally clear, clinically relevant, and well-motivated. It successfully frames why automated TUG segmentation is important and identifies the clinical gap addressed by the paper. I have no comments here.

Methodology. In the “Participants” section, please justify the number of participants. Figure 1 – The upper part with the personalities is not very legible. Please change it.

How did you calculate entropy?

How did you calculate signal energy?

Please increase the font size in the figures. Figure 6 – please add an explanation of the abbreviations for the vertical axis.

Author Response

Dear Reviewer, thank you for taking the time to provide a detailed review of our paper. A summary of our responses to your suggestions is provided in the attached file.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Authors have substantially addressed the comments made to the previous version and improved the language too.

Author Response

Comments and Suggestions: The authors have substantially addressed the comments made to the previous version and improved the language too.
Dear Reviewer, thank you very much for your time and effort to deeply review the paper. 
Your comments really guided us to improve the quality of the paper up to the level of the 
journal.