A Lightweight Convolutional Neural Network Method for Two-Dimensional PhotoPlethysmoGraphy Signals

Round 1

Reviewer 1 Report

This paper aims to apply PPG to biometric authentication in wearable devices using lightweight machine learning methods. The paper is a useful study for the future because it achieves high accuracy by 2D transformation to PPG signals and lightweight by applying LW-CNN. On the other hand, However, I think that there are several points in this paper that lack explanation, need to be corrected, or are questionable:

Major revision:

1)     In the introduction, the previous studies are described, but the problems in the previous studies and the novelty and position of this study should be described.

2)     In the third paragraph of section 2.3 (page 4), it is written "eight-minute PPG signal", does this authentication require a long PPG signal? We think that this time is a big problem in personal authentication. Please discuss and describe the limitations of this method and how it can be improved.

3)     The second paragraph of section 2.1 says that the dynamic range of the PPG signal is unified to 1, but looking at Figure 2 (left), this is not the case. please explain or change to the correct one for the PPG signal in Figure 2.

4)     In the third paragraph of section 2.3 (page 4), it is written "two-dimensional images with a size of 28×28", but Figure 2 (right) is smaller than that. Which is correct ?

5)     In the first paragraph of section 3.1, the word "We carefully selected PPG signals from 16 individuals as representative and valid research samples to ensure precise experimental analysis." is used, but the criteria for selection should be detailed. Also, if possible, the mean and standard deviation of the age and the percentage of gender should be given for the selected sample.

6)     Is it possible to compare the parameters and training time between the proposed method and methods using 1D PPG signals (1D-CNN and LSTM)? If possible, the comparison results should also be described.

Minor revision:

1)     In the first paragraph of section 2. 1 (page 3), there is a description of the noise that occurs in the PPG signal, but a reference should be added.

2)     What are the reasons for using a 3rd-order Butterworth filter among the various types of filters ? (Second paragraph of section 2. 1, page 4)

3)     Explain what N, X , and q represent in equation (2).

4)     In figure 2(right), a color bar should be added so that the values represented by the colors can be seen.

5)     The first paragraph of chapter 2.3 (page 5) describes LW-CNN, but a reference should be added.

6)     The title of chapter 3.1 is "Datastest", but "Dataset" is correct.

7)     Figure 8 and Figure 9 use bar charts, but tables are easier to understand.

Author Response

Dear reviewer,

Thank you very much for reading my paper and giving your comments.

Regarding your major revision 1, at the end of the first section, I added questions about the existing problems in the previous literature and the novelty of my research.

With regard to your major revision 2, in the third paragraph of section 2.3, I re-revised the presentation about my data.

Regarding your major revisions 3 and 4, I have re-modified the images on the left and right of Figure 2.

With regard to your major revision 5, in Section 3.1, I have explained the composition of my individual selection, including gender and age.

Regarding your major revision 6, I have described the parameters and training time of 1d-cnn and lstm in Table 2.

Regarding your minor revision 1, I made a reference for the noise sources of ppg signals.

Regarding your minor revision 2, in the second paragraph of Section 2.1 I explained in detail the reasons why the third-order Butterworth bandpass filter and its cutoff frequency were chosen.

With regard to your minor revision 3, in Section 2.2 (4), I reexplained what the letters stand for.

For your little revision 4, on the right of Figure 2, I added a color bar.

As for your minor revision 5, LW-CNN is a deep learning model designed by me, and the methods used on it include deep separable convolution and residual join I have added references.

Regarding your minor revision 6, I have changed the title of Section 3.1.

For your minor revision 7, I have replaced figures 8 and 9 with Tables 1 and 3 in the present article.

Attached is my revised paper. Thank you again for reviewing my paper.

Sincerely,
Zhang Xudong

Reviewer 2 Report

This paper developed an new identity authentication method by using a LW-CNN method for PPG signals. Introduction must be improved for publication, It is difficult to find connections between wearable device and proposed method. But proposed method has a novelty in LW-CNN design. 

Specific comments. 

1) In line (2~5) of page 2, you can explain more clearly why you adapt PPG authentication instead of other methods. I recommend continuous user authentication as a reason.

2) You can compare 1D and 2D CNN in the aspect of real time implementation.

3) You can explain more clearly the difference between W of Eq(1) and M of Eq(2).

5) In figure 3, you can add more information(kernel, receptive field area...etc) to describe Depthwise Separable Convolution. 

6) In figure 5, you can describe '+' location of residual connection more clearly.

8) You can add the references in 1D-CNN and LSTM in page 9.

9) In table 1, you can add information if AlexNet and GoogleNet are trained from initial or transfered by replacing top layers. 

10) It is difficult to find 98.62% and 96.17% are good results. You can compare this methods with online PPG or other realtime methods. 

4) You can check the grammar in line (6,7) of page 6. (x is the previous The output of one or more layers,)

7) You can check the grammar in line 3 of page 9. (Precision rate, precision rate, recall rate, and F1 value can also be calculated by analyzing the confusion matrix.)

Author Response

Dear reviewer,

Thank you very much for reading my paper and giving your comments.

Regarding your suggestion 1, I have added the reason why ppg signals can be used for identity authentication on lines 1 to 6 of the second page.

As for your suggestions 2 and 10, my authentication method is based on batch processing. In my comparative experiments, it is obvious that it has advantages over traditional one-dimensional timing signal and two-dimensional classical deep learning model.

Regarding your suggestion 3, I have revised the relationship between these two formulas in paragraph (4) of Section 2.2.

Regarding your suggestions 5 and 6, I have revised my two drawings. Figure 3 I added a description of the convolution kernel. In Figure 5, I added a note for the "+" of the residual connection.

For your recommendation 8, I added references about 1d-cnn and lstm.

For your recommendation 9, I have added information from the initial layer training in Tables 3 and 4.

Regarding your suggestions 4 and 7, I have corrected the grammatical errors in my article.

Attached is my revised paper. Thank you again for reviewing my paper.

Sincerely,
Zhang Xudong

Reviewer 3 Report

The paper is well-written and presents a novel approach. Overall, while the paper presents a promising approach, addressing these potential drawbacks or weaknesses could further strengthen the work and provide a more comprehensive understanding of the method's limitations and practical implications:

1. Dataset limitations. The experiments were conducted using PPG signals from only 16 individuals from the BIDMC dataset. While this provides a good proof-of-concept, a larger and more diverse dataset would be needed to thoroughly evaluate the method's generalizability and robustness across different populations and conditions.
2. Signal quality and noise. The paper briefly mentions the preprocessing steps to remove noise from the PPG signals, but it does not delve into the specifics of how the method handles varying signal quality or significant noise levels. In real-world scenarios, PPG signals can be susceptible to motion artifacts, environmental factors, and other sources of noise, which could affect the performance of the proposed method.
3. Real-time performance evaluation. The paper focuses on the accuracy and efficiency of the proposed method, but it does not evaluate the real-time performance or latency of the authentication process. In wearable device applications, real-time performance and low latency are crucial factors that could impact the user experience and practicality of the method.

Author Response

Dear reviewer,

Thank you very much for reading my paper and giving your comments.

Regarding your comment 1, dataset limitation, my dataset is indeed a relatively small dataset. However, the data selected in my dataset are very representative, and I make sure that gender and all age groups are present in my dataset. In Section 3.1 of the paper, I revised the introduction to data sets. But expanding the data set is a little difficult due to my device limitations.

In response to your comment 2, I have revised my content in Section 2.1 to reintroduce the filtering range of my choice of a third-order Butterworth bandpass filter, which filters out the common noise of PPG signals.

Regarding your comment 3, I have introduced the total number of parameters， total parameter size and iteration time per iteration of the neural network I selected in Table 2 and Table 4. It can be seen that my scheme is suitable for resource-constrained environments such as wearable devices.

Attached is my revised paper. Thank you again for reviewing my paper.

Sincerely,

Zhang Xudong

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

The authors have satisfactorily complied the review comments in the revised manuscript, thank you. The revised manuscript is much improved. However, there is a point in this paper that needs to be corrected:

1) In Figure 2(right), the horizontal and vertical axes are not equally spaced, and the width of 15-20 is wider than that of 20-28, even though the width of 20-28 should be wider than that of 15-20.

Author Response

Dear reviewer,

Thank you again for reviewing my paper and giving your comments.

I have modified the problem with the axes in Figure 2 (right). Make sure 20-28 is wider than the rest. Thank you for your reminding.

Attached is my revised manuscript.

Sincerely,
Xudong Zhang

Reviewer 2 Report

Authors responded to all my comments. 

But I have one additional question in your revised manuscript. 

(In Table 2) 236834/44016 =5.38; 2.68/1.47 = 1.82

(In Table 4)1048784 / 44016 = 23.83; 4.49/ 1.47 = 3.05

Why ratios between "Total Number of Parameters" and "Total Parameter Size" are presented differently in your tables?

Author Response

Dear reviewer,

Thank you again for reviewing my paper and giving your comments.

I have re-modified the data of the total size of the parameters in my Table 2 and Table 4. Thank you for your reminding.

Attached is my revised manuscript.

Sincerely,
Xudong Zhang

Round 3

Reviewer 2 Report

Authors responded to all my comments properly. The additional question corrected errors of presentation properly.