Research on sEMG Feature Generation and Classification Performance Based on EBGAN

Round 1

Reviewer 1 Report

1. The state of the art must refer basically also to two main problems according to authors claim as contribution: energy generative adversarial network and buildings using scarce data. There are no reference in this direction, but there are in the literature. Few examples are:

a) Junbo Zhao, Michael Mathieu and Yann LeCun, ENERGY-BASED GENERATIVE ADVERSARIAL NETWORKS, Published as a conference paper at ICLR 2017

b) Gaby Baasch, Guillaume Rousseau, Ralph Evins, A Conditional Generative Adversarial Network for energy use in multiple buildings using scarce data, Energy and AI, 2021

2. The new approaches must clearly explain, e.g., "4.3.1. Generate sample data quality and generalization ability". The "generalization ability" must be clearly explained with examples.

3. What means "cheat discriminator"?

4. Nash equilibrium (game theory) must be explained in this case of DLNN.

5. The energy form from Fig. 5 is not presented.

6. The formula from line 216 cannot be read.

7. The maximum mean discrepancy (MMD) [18] must be clear presented as measure between e two distributions. In biomedical signals (e.g., EMG), the similarity between signals is not always valid. The authors must clearly explain (graphical) how the EMG that uses only distribution can be useful for clinical practice.

8. The originality of the paper must be clearly presented. 

Author Response

Dear reviewers:

We thank the experts for their suggestions. These suggestions helped to improve the quality of the manuscript! We revised each suggestion and improved the manuscript accordingly, showing changes and additions in red in the manuscript. The full text of the manuscript was revised and edited by native English speakers.

Expert 1:

1. The state of the art must refer basically also to two main problems according to authors claim as contribution: energy generative adversarial network and buildings using scarce data. There are no reference in this direction, but there are in the literature. Few examples are:
2. a) Junbo Zhao, Michael Mathieu and Yann LeCun, ENERGY-BASED GENERATIVE ADVERSARIAL NETWORKS, Published as a conference paper at ICLR 2017
3. b) Gaby Baasch, Guillaume Rousseau, Ralph Evins, A Conditional Generative Adversarial Network for energy use in multiple buildings using scarce data, Energy and AI, 2021

Thank you for your suggestions. These papers were cited in the manuscript and added to the references. Please check the revised references [8] and [15] for details.

2. The new approaches must clearly explain, e.g., "4.3.1. Generate sample data quality and generalization ability". The "generalization ability" must be clearly explained with examples.

Thank you for your comments. In this section, five models were used to classify and identify the original dataset and the synthetic dataset, and the recognition accuracy of the five models for the synthetic dataset was improved compared to the original dataset. In response to the expert’s opinions, we feel that it is more appropriate to change to "universality" for the adaptation of various models. Please check the heading of Section 4.3.1 in the revised manuscript.

3. What means "cheat discriminator"?

Thank you for your question. This is a description error. It means that the generator constantly inputs the generated high-quality samples into the discriminator to confuse the discriminator's judgment and increase the possibility of the generated samples being recognized as real samples. We corrected this description.

4. Nash equilibrium (game theory) must be explained in this case of DLNN.

Thank you for your comments. Nash equilibrium (game theory) has been explained in the literature related to generative adversarial networks, and we added the corresponding literature citations, adding a description of Nash equilibrium in the revised draft; please refer to lines 190–191. The literature cited is as follows:

[15] Zhao J, Mathieu M, Lecun Y. Energy-based generative adversarial network[J]. arXiv preprint,2016, arXiv:1609.03126

5. The energy form from Fig. 5 is not presented.

Thank you for your suggestions. The energy form of discriminator is the loss function of the discriminator , which is expressed by Formula (9), , and the corresponding description of energy form was added to the revised manuscript.

6. The formula from line 216 cannot be read.

Thank you for your comments. This could be due to an incompatibility with the PDF format. The formula in line 216 is: , which is an interpretation of Equation 9. Please see line 228 of the revised manuscript for details.

7. The maximum mean discrepancy (MMD) [18] must be clear presented as measure between e two distributions. In biomedical signals (e.g., EMG), the similarity between signals is not always valid. The authors must clearly explain (graphical) how the EMG that uses only distribution can be useful for clinical practice.

Thank you for your comments.

On the one hand, related fields have used this indicator to evaluate the validity of time-series data, such as:

[1] Xu Q, Huang G, Yuan Y, et al. An empirical study on evaluation metrics of generative adversarial networks[J]. arXiv Preprint, 2018, arXiv:1806.07755

[2] Jin Q, Lin R, Yang F. E-WACGAN: enhanced generative model of signaling data based on WGAN-GP and ACGAN[J]. IEEE Systems Journal, 2019, 14(3): 3289-3300.

We referred to the methods of the above-cited literature and used MMD to evaluate the similarity of the time-domain features of EMG signals.

On the other hand, since the time-domain characteristics of EMG signals are extracted to describe the strength of signals in the time series, we believe that it is applicable and feasible to use MMD to evaluate time series similarity.

8. The originality of the paper must be clearly presented. 

Thank you for your comments. The contribution and originality of this article are explained in the Introduction. Please see lines 52–64 of the revised version for details.

Author Response File: Author Response.docx

Reviewer 2 Report

This paper proposed a generative adversarial network that generates synthetic surface electromyography signals. This GAN is developed based on time-domain features from surface electromyography signals. The idea of this paper is interesting, but the manuscript needs some revision before accepting for publishing.

1. The authors mentioned classification several times and presented the classification accuracy. However, I did not see description on what is classified. It is not clear to me what exactly the classification task does.

2. The meanings of some sentences are not clearly, and some sentences have grammar errors. 

For example, the long sentence in line 45-49 contains three clauses. I believe the first clause says the "sharing data" will lead to privacy issue but not the "difficult to achieve data sharing" will lead to this issue. The "difficult to share data" will lead to lack of dataset but not the "privacy problems". I suggest to break the long sentences into shorter ones so that the meaning will not alter. Similar long sentences are in line 75-77 and line 92-95.

Line 103-105 talks about the advantage of invasive acquisition, I'm not sure why authors conclude that it is not suitable for acceptance.

Line 141-145 uses "because... so...".

In paragraph 149-156, I did not get why time domain features are better than frequency domain features.

3. In section of "Related Work", I would like to see some example for ML-based enhancement algorithms since the term ML is too broad here.

4. In line 216, the equation misses some symbols.

5. I did not find Table 2 (mentioned in line 225).

6. In classification experiments (Table 4), I did not find the description for testset. Was it obtained from splitting the whole dataset by 20% or it is obtained from other sources?

7. In Section 4.3.2, I did not get what line 339-340 tries to convey. Is the mixed set is in size of 0.5-3 times than the original dataset? If it is the case, what's the ratio of the synthetic data?

8. I would like to see how the performance comparison between the model trained on syntheitc dataset and test on the real dataset, and the model trained on real dataset and test on real dataset. This may be a more direct evidence on how good the synthetic dataset is.

Author Response

1. The authors mentioned classification several times and presented the classification accuracy. However, I did not see description on what is classified. It is not clear to me what exactly the classification task does.

Thank you for your comments. The classification task is to classify the five lower limb movements, which means using walking on the flat ground, going upstairs, going downstairs, sitting down and standing up as the classification objects, and the corresponding descriptions were added in Section 4 in the revised manuscript. Please check lines 299–301 in Section 4.3 for details.

2. The meanings of some sentences are not clearly, and some sentences have grammar errors. 

For example, the long sentence in line 45-49 contains three clauses. I believe the first clause says the "sharing data" will lead to privacy issue but not the "difficult to achieve data sharing" will lead to this issue. The "difficult to share data" will lead to lack of dataset but not the "privacy problems". I suggest to break the long sentences into shorter ones so that the meaning will not alter. Similar long sentences are in line 75-77 and line 92-95.

Line 103-105 talks about the advantage of invasive acquisition, I'm not sure why authors conclude that it is not suitable for acceptance.

Line 141-145 uses "because... so...".

In paragraph 149-156, I did not get why time domain features are better than frequency domain features.

Thank you for your kind suggestions.

For lines 45–49, 75–77 and 93–98, relevant grammatical errors were corrected in the revised manuscript. Please see lines 45–49, 75–77 and 93–98 of the revised manuscript for details.

For lines 103–105, in the invasive collection method, the needle needs to be inserted into the muscle, which will damage the muscle, and, as the collected material is small and thus cannot reflect the characteristics of the entire muscle, non-invasive collection was adopted. The reason was added in the revised manuscript; please check lines 109–111.

In lines 149–156, as time domain features can reflect the change in signal strength with time, without other transformation processing, are easy to obtain and have a lower computational load, it was more appropriate to select time domain features in this paper.

3. In section of "Related Work", I would like to see some example for ML-based enhancement algorithms since the term ML is too broad here.

Thank you for your comments. Relevant examples of enhancement algorithms based on machine learning were added to the article, such as synthetic oversampling SMOTE ^[9] and adaptive synthetic sampling ADASYN ^[10]. See lines 69–70 of the revised version for details.

[9] Chawla N V , Bowyer K W , Hall L O , et al. SMOTE: synthetic minority over-sampling technique[J]. AI Access Foundation, 2002(1).

[10] He H , Yang B , Garcia E A , et al. ADASYN: Adaptive synthetic sampling approach for imbalanced learning[C]// Neural Networks, 2008. IJCNN 2008. (IEEE World Congress on Computational Intelligence). IEEE International Joint Conference on. IEEE, 2008.

4. In line 216, the equation misses some symbols.

Thank you for your comments. This could be due to an incompatibility with the PDF format. The formula in  line 216 is: , which is an interpretation of Equation 9. Please see line 228 of the revised manuscript for details.

5. I did not find Table 2 (mentioned in line 225).

Thank you for your comments. Table 2 was added to the revised manuscript. Please see line 241 in Section 3.2 of the revised version for details.

6. In classification experiments (Table 4), I did not find the description for testset. Was it obtained from splitting the whole dataset by 20% or it is obtained from other sources?

Thank you for your comments. It is obtained by splitting 20% of the dataset, which was explained in the text. See lines 315–316 of the revised version for details.

7. In Section 4.3.2, I did not get what line 339-340 tries to convey. Is the mixed set is in size of 0.5-3 times than the original dataset? If it is the case, what's the ratio of the synthetic data?

Thank you for your comments. It means that the scale of the newly generated data in the mixed set is 0.5~3 times that of the original data set, so the synthetic data ratios in differently mixed sets are 33%, 50%, 66% and 75%. The corresponding description was added in the revised manuscript; please check lines 357–359.

8. I would like to see how the performance comparison between the model trained on syntheitc dataset and test on the real dataset, and the model trained on real dataset and test on real dataset. This may be a more direct evidence on how good the synthetic dataset is.

Thank you for your suggestion, which is very informative and thus we will consider applying it in subsequent research. However, in the current research work, we used the common comparison method of similar research and obtained relatively ideal results; thank you for your understanding.

Author Response File: Author Response.docx

Reviewer 3 Report

This manuscript reports a study where the authors have proposed a new feature generation method based on energy generative adversarial network (EBGAN) for improving the accuracy of sEMG classification. The results of the tests that were carried out are promising.

Overall, the manuscript is excellently written and attractive to read. The background and the contribution are well explained in the introduction section. The methods and procedures are explained, and the results are presented understandably.  

In table 1, where the basic testers are presented, we can see data about testers' height and weight. However, the text did not explain why these parameters are essential. Are these important at all?

In figure 8, I suggest explaining which data/line corresponds to which losses (generators vs discriminator losses). In its current form, it needs to be clarified.

In conclusion, I suggest adding a discussion on the study's main limitations.

Author Response

1、In table 1, where the basic testers are presented, we can see data about testers' height and weight. However, the text did not explain why these parameters are essential. Are these important at all?

Thank you for your comments. The data in Table 1 objectively describe the situation of the subjects in this article, which is a basic work for the readers' reference.

2、In figure 8, I suggest explaining which data/line corresponds to which losses (generators vs discriminator losses). In its current form, it needs to be clarified.

Thank you for your suggestion. The generator and discriminator losses were indicated in Figure 8 of the revised manuscript.

3、In conclusion, I suggest adding a discussion on the study's main limitations.

Thank you for your suggestion. A discussion of the limitations on the study was added to the Conclusion of this paper. Please check lines 410–417 of the revised version for details.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

All the requirements are fulfilled. The paper can be published now. 

Author Response

Dear reviewers:

We thank the experts for their suggestions. These suggestions helped to improve the quality of the manuscript! We revised each suggestion and improved the manuscript accordingly, showing changes and additions in red in the manuscript. The full text of the manuscript was revised and edited by native English speakers.

Reviewer 2 Report

The authors have addressed all my concerns. I appreciate their efforts. 

The only minor revision needed is for formatting issues in:

1) line 190: there is an empty space in line

2) line 204: looks like there is part of sentences needs to be removed but the symble is still there

3) line 360-361: This sentence should be in the same line.

Author Response

Dear reviewers:

We thank the experts for their suggestions. These suggestions helped to improve the quality of the manuscript! We revised each suggestion and improved the manuscript accordingly, showing changes and additions in red in the manuscript. The full text of the manuscript was revised and edited by native English speakers.

Expert 2:

1.  line 190: there is an empty space in line:

Thank you for your comments. The empty space in line 190 has been removed.

2.  line 204: looks like there is part of sentences needs to be removed but the symble is still there.

Thank you for your comments. The symbol in line 204 has been removed.

3. line 360-361: This sentence should be in the same line.

Thank you for your suggestions. The sentence in line 360-361 has been edited to the same line.

Author Response File: Author Response.pdf