PIFall: A Pressure Insole-Based Fall Detection System for the Elderly Using ResNet3D
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors provide a method of measuring fall detection for the elderly utilising a ResNet3D algorithm. Several issues should be addressed before the paper is published:
1) It would help to define the novelty of the method given the number of papers that are available in this research area.
2) The number of samples in the paper is given as 65. How was this number of samples was chosen? The authors should provide examples the data sets with different numbers of samples to show this was the optimum number.
3) Only data from one human subject appears to be presented. There need to be enough subjects to show the statical relevance.
4) The references are not up to date, see: “Smart Insole Monitoring System for Fall Detection and Bad Plantar Pressure, March 2022, DOI: 10.1007/978-3-030-99619-2_20. Subramaniam S, Faisal AI, Deen MJ. Wearable Sensor Systems for Fall Risk Assessment: A Review. Front Digit Health. 2022 Jul 14;4:921506. doi: 10.3389/fdgth.2022.921506. PMID: 35911615; PMCID: PMC9329588. Mun, F., Choi, A. Deep learning approach to estimate foot pressure distribution in walking with the application for a cost-effective insole system. J NeuroEngineering Rehabil 19, 4 (2022). https://doi.org/10.1186/s12984-022-00987-8. The authors need to review the literature full as there are papers that used similar approached to their method.
Author Response
Please see the attachment
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper presents a fall detection system with future potential for use by the elderly. The article is well written, although there is some redundancy of information between some sections of the paper.
In my review I make the following questions and/or annotations that should be answered by the authors, in my opinion, to improve the quality of the paper.
How do the possible problems that older people, who due to some injury, do not walk in a normal way affect the detection? For this same reason, it may be that the method presented requires, to be effective, a personalized training for each particular case. What is the authors' opinion in this regard?
Lines 201 to 209 contain text that has appeared before.
The diagram of the physical circuit seems to indicate that it is of considerable size to be worn in shoes. I understand that it is a prototype that still requires a large future development to be improved.
Section 4.3.2 describes how the pressure sensors are calibrated for different people. This section should be expanded as it appears to be critical to the proper functioning of the system. Subtracting each sensor from its mean value when it has no load seems like a good idea, but it is somewhat difficult to do (i.e., how do you know accurately that the sensors have no load? The foot motion is quite complex, which can make this process difficult to implement.
I don't understand what function is being described in the text on lines 269 to 274. This should be better explained.
The text on lines 280 to 281 does not seem to make sense.
In lines 282 and 283 it is written "The traditional ResNet model is not effective in classifying video data. So in this study, we utilize a variant of ResNet, namely ResNet(2+1)D". In this paper we are not processing video data, so I do not understand the premise or the consequence of these lines.
The period in line 302 should be a comma.
Lines 414 to 416 state "During fall actions, to enhance model stability and ensure safety, participants are allowed to fall in any position except the direction of the fall, without any specific requirements." The authors should clarify what exactly this restriction means.
I agree with the authors that there are few experiments to effectively train an ANN. On the other hand, it seems that all available patterns have been used to perform the training, and this may be a problem. Proper model evaluation requires that the confusion matrix be created for a set of patterns that have not been used to train the ANN. Authors should explain the training process in more details.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe authors propose the use of a pressure insole-based fall detection system for the elderly, utilising the ResNet3D algorithm. I have several observations that should be addressed.
1) There is no indication as to how changes in temperature of the insole material when the patient is walking are addressed, as this will influence the pressure sensor conductance.
2) What is the resolution of the insole measurements, mm/cm? As this will affect the accuracy of the measurements. Also, what is a sampling rate of the measurements, i.e. temporal resolution? This should be made clear.
3) The statement” The system achieves an overall accuracy of 91% in detecting falls and 94% in identifying specific 12 fall actions”. However, it is not clear what is used as a gold standard to results are validated against.
4) Given there are methods to print conductive tracks on fabric the use of stitching does not seem to be an accurate and time-consuming method to construct the insole.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
