Dual Head and Dual Attention in Deep Learning for End-to-End EEG Motor Imagery Classification

Round 1

Reviewer 1 Report

The paper "Dual Head and Dual attention in Deep Learning for End-to-End EEG Motor Imagery Classification" proposes to improve feature selection of MI-EEG signals using Deep Learning. The authors develop a new neural network architecture called DHDANet mainly focused on feature extraction in the temporal and spatial spectrum using dual attention mechanisms in order to help increase the classification performance of MI-EEG signals.
The results of the paper are interesting, as well as the solution proposed by the authors. However, there are several errors that need to be resolved in order to reconsider the acceptance of the work:

• Abstract: Change "setts" to "set".
• Line 20:  Change "0. Introduction" to "1. Introduction" (should start with 1).
• Line 369: Removes the sign "=".
• Line 409 : Non-existence of cross-referencing -> "experimental results shown in Tab.??".
• Line 425 : Non-existence of cross-referencing -> "The experimental results of these frameworks are shown in Fig.??".
• Line 426 : Non-existence of cross-referencing -> "results of classification accuracy are shown in Tab.??".
• Line 446:  "The the" -> Removes "the".
• Lines 462 and 463: hash search is duplicated -> "Hash search and hash search are two methods".
• Figure 8: exceed the margin limits.
• Figure 9: should be resized.
• All figures should be scalable vector graphics (e.g. svg or eps) instead of rasterized images.
• An space when referencing authors have to be inserted (before brackets).
• References about "overfitting" should be added.
• Do not use abbreviations for the words Tables, Figures, or Equations that correspond to cross-references. Also, add a space after the corresponding reference number.
• Add details about the configuration of the computer (hardware and software) used for experiments. 

On the other hand, there is no statistical analysis of the results to determine whether the differences in accuracy were statistically significant. Please add some statistical analysis. Moreover, English should be checked carefully. There are some grammatical errors, which, however, do not impede the understanding of the manuscript.

Author Response

We thank you for offering the valuable comments and suggestions, which helped us to improve the quality of the paper. We apologize for these mistakes .The following answers addressed your comments.

 

Point 1: Abstract: Change "setts" to "set". 


Response 1: It’s corrected, see line 8.

 

Point 2: Line 20:  Change "0. Introduction" to "1. Introduction" (should start with 1). 
 

Response 2: It’s corrected.

 

Point 3: Line 369: Removes the sign "=". 
 

Response 3: It’s corrected, see line 349 in the new edition of the paper.

 

Point 4: Line 409 : Non-existence of cross-referencing -> "experimental results shown in Tab.??". 
 

Response 4: It’s corrected, see line 390.

 

Point 5: Line 425 : Non-existence of cross-referencing -> "The experimental results of these frameworks are shown in Fig.??". 
 

Response 5: It’s corrected, see line 406.

 

Point 6: Line 426 : Non-existence of cross-referencing -> "results of classification accuracy are shown in Tab.??". 
 

Response 6: It’s corrected, see line 407.

 

Point 7: Line 446:  "The the" -> Removes "the". 
 

Response 7: It’s corrected, see line 430

 

Point 8: Lines 462 and 463: hash search is duplicated -> "Hash search and hash search are two methods". 
 

Response 8: "Hash search and hash search" has been modified to "Hash search and bash search", see line 447.

 

Point 9: Figure 8: exceed the margin limits. 
 

Response 9: It has been resized.

 

Point 10: Figure 9: should be resized. 
 

Response 10: It has been resized.

 

Point 11: All figures should be scalable vector graphics (e.g. svg or eps) instead of rasterized image.
 

Response 11:  The images are used Python matplotlib.pyplot to generate the experiment results to PNG images. We  would  like  to  thank  for  raising  this  concern. I'm going to explore how do you generate vector graphics.

 

Point 12: An space when referencing authors have to be inserted (before brackets). 
 

Response 12: Thank  you very much  for  your  comments. We have checked and corrected this problem thoroughly.

 

Point 13: References about "overfitting" should be added. 
 

Response 13: See line 282, we added two references, as followed:

[1]Vincent, P.; Larochelle, H.; Lajoie, I.; Bengio, Y.; Manzagol, P.A. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of machine learning research2010,11, 3371–3408.

[2]Pang, G.; Shen, C.; Cao, L.; Hengel, A.V.D. Deep learning for anomaly detection: A review. ACM Computing Surveys (CSUR)2021,54, 1–38

 

Point 14: Do not use abbreviations for the words Tables, Figures, or Equations that correspond to cross-references. Also, add a space after the corresponding reference number. 
 

Response 14: The whole paper has been improved according to this requirement.

 

Point 15: Add details about the configuration of the computer (hardware and software) used for experiments. 
 

Response 15: We added the description of the experimental operating environment at line 332 of paper. After the paper is published, we will open the source code to github.

 

Point 16: On the other hand, there is no statistical analysis of the results to determine whether the differences in accuracy were statistically significant. Please add some statistical analysis. 
 

Response 16: We improved this analysis, see lines 392 to 396:

Moreover, compared with the other four algorithms, DHDANet has a high recall rate, indicating that the recognition rate of poor samples is satisfactory. Meanwhile, the speciality is also high, indicating a low false positive rate for samples with non-motor imagery. DHDANet has high sensitivity and specificity for MI recognition which confirms the superiority and is a popular choice for high performance diagnostic.

 

Point 17: Moreover, English should be checked carefully. There are some grammatical errors, which, however, do not impede the understanding of the manuscript. 
 

Response 17: Thank you very much for your remind. We did our best to check the grammar and improve the description.

Reviewer 2 Report

The authors provided well organized manuscript on an interesting topic. In my opinion, it suits for publication after minor improvements.

Noticed issues:

• The introduction should be not numbered;
• Tables and figures number in lines 409, 425, 426 are not specified

In figure 12, provided results show a very huge range of error values, almost equal to the main value. Such results must be additionally discussed.

Author Response

We thank you for offering the valuable comments and suggestions, which helped us to improve the quality of the paper. The following answers addressed your comments.

Point 1: The introduction should be not numbered. 
 

Response 1: We change "0. Introduction" to "1. Introduction", see line 20.

 

Point 2: Tables and figures number in lines 409, 425, 426 are not specified.

Response 2: We apologize for this mistakes. Them has been corrected. Please see lines 390, 406,407 in the revised paper.

 

Point 3: In figure 12, provided results show a very huge range of error values, almost equal to the main value. Such results must be additionally discussed.

Response3: Because there is many more of clutter for the not MI illiteracy subjects to do the MI experiment. It indicates the superiority of DHDANet. please see the detail discussion in lines between 425 to 432.

Reviewer 3 Report

This study proposes an electroencephalogram (EEG) motor imagery recognition system for people with severe motor impairments. The approach is interesting for BCI researchers and the audience of Sensors. The manuscript is easy readable, but it should be organized better to facilitate the reader understanding. Furthermore, it can be benefited after proofreading by a native English speaker. For a better comparison with the state of the art, other well-known datasets, such as BCI Competition IV datasets 2a and 2b, and BCI Competition III dataset IVa. I have some comments and suggestions, as follows:

Introduction

1- In section Introduction, the authors should comment other recent/relevant references, such as: 

1.1) Zuo, Cili, et al. "Cluster decomposing and multi-objective optimization based-ensemble learning framework for motor imagery-based brain–computer interfaces." Journal of Neural Engineering 18.2 (2021): 026018.

1.2) Chu, Yaqi, et al. "Decoding multiclass motor imagery EEG from the same upper limb by combining Riemannian geometry features and partial least squares regression." Journal of Neural Engineering 17.4 (2020): 046029.

1.3) Barachant, Alexandre, et al. "Multiclass brain–computer interface classification by Riemannian geometry." IEEE Transactions on Biomedical Engineering 59.4 (2011): 920-928.

1.4) Yger, Florian, Maxime Berar, and Fabien Lotte. "Riemannian approaches in brain-computer interfaces: a review." IEEE Transactions on Neural Systems and Rehabilitation Engineering 25.10 (2016): 1753-1762.

1.5) D. Milanés Hermosilla et al., "Shallow Convolutional Network Excel for Classifying Motor Imagery EEG in BCI Applications," in IEEE Access, vol. 9, pp. 98275-98286, 2021, doi: 10.1109/ACCESS.2021.3091399.

2- In addition to clinical BCI studies that used ERD/ERS for motor rehabilitation, also comment some review and research that demonstrate the potential of using ERD/ERS patterns. Please see some works:

2.1) Pfurtscheller, Gert, and FH Lopes Da Silva. "Event-related EEG/MEG synchronization and desynchronization: basic principles." Clinical neurophysiology 110.11 (1999): 1842-1857.

2.2) Hashimoto, Yasunari, and Junichi Ushiba. "EEG-based classification of imaginary left and right foot movements using beta rebound." Clinical neurophysiology 124.11 (2013): 2153-2160.

2.3) D. Milanés Hermosilla et al., "Shallow Convolutional Network Excel for Classifying Motor Imagery EEG in BCI Applications," in IEEE Access, vol. 9, pp. 98275-98286, 2021, doi: 10.1109/ACCESS.2021.3091399.

2.4) Graimann, Bernhard, et al. "Visualization of significant ERD/ERS patterns in multichannel EEG and ECoG data." Clinical neurophysiology 113.1 (2002): 43-47.

3- Other studies should be referenced on the lines 79-81, where the authors comment the following idea: This is a common problem with complex CNN models which result in more training time and worse real-time performance. Please discuss about the BCI latency permissible to users to feel a real closed-loop, and consequently enhance neuroplasticity. See some related studies:

3.1) Romero-Laiseca, Maria Alejandra, et al. "A low-cost lower-limb brain-machine interface triggered by pedaling motor imagery for post-stroke patients rehabilitation." IEEE Transactions on Neural Systems and Rehabilitation Engineering 28.4 (2020): 988-996.

3.2) Mrachacz-Kersting, Natalie, et al. "Efficient neuroplasticity induction in chronic stroke patients by an associative brain-computer interface." Journal of neurophysiology 115.3 (2016): 1410-1421.

1. Data

4- In section 1. Data, specify the frequency range used to acquire the EEG data. The authors should justify why 20 EEG channels were used over the referred locations. Please cite similar motor imagery classification studies that also used these locations.

2. Method

5- In general, this section 2. Method should be better organized to facilitate the reader understanding, and avoid redundancy. For instance, the authors can first give an overview about the proposed system, using Figure 4 and part of the text in section 2.3. DHDANet The authors combine here present tense and past tense. I suggest to use only present tense describing the proposed recognition system. Also Figures 5 and 6, and their corresponding texts can be moved to sections 2.2 to avoid the related information disperse, and consequently facilitate the reader understanding. Notice that some variables are not defined, and sometimes other variables are defined too far of the first mention. Please define all variables, preferably close to the first mention. Also avoid to use the same variable for different meanings.

6- Check in the subsection 2.1 Input Data the following phrases (see lines 163 and 170): "E is the number of epochs and 20 MI areas epochs"; "....of E-recorded electrodes". It is confusing. Also check the sentence on the lines 171-172 pp. 5.

7- The authors affirms that the crops were collected starting on the trial cue (see lines 166 pp. 5). It is recommend to the the first 500 ms after beginning the cue due to event-related potentials (ERPs) that are spontaneously generated in the brain, and can benefit the BCI performance, although the classification output could not be correlated (or related) with the user's motor intention (MI tasks). Please see

7.1) Wolpaw J R and Boulay C B 2009 Brain signals for brain– computer interfaces Brain–Computer Interfaces (New York: Springer) pp 29–46

7.2) Pfurtscheller G and Neuper C 2009 Dynamics of sensorimotor oscillations in a motor task Brain–Computer Interfaces (New York: Springer) pp 47–64 

8- In Eq. (1) define i and j, X1 and X2, d2. Also move Figure 4 to its first mention. Make a similar action with each figure throughout all text. The quality of Figure 4 should be improved. Also edit this figure to correct some texts.

9- Replace "T·E" by "T×E" on the line 169. Also replace "Equ" by "Eq." on the line 228. Make a similar action throughout all text.  Please define i, p and q, B, and D in subsection 2.2.1 (see lines 217-231). Make a similar action with i, j, and k on the line 253.

10- Cite studies that also support the following affirmation (see lines 282-283): "the time period for an ERD/ERS....... between 500 ms to 1 s".

3. Experiments and Results

11- Various redundant texts should be removed in this section 3. Experiments and Results. It is one of the main problems found in this section. See here some redundant text that can be removed: lines 317-323, lines 332-364 in subsection 3.1, lines 375-380

12- Add other MI datasets for better comparison with the state-of-the-art. Also include a time consuming analysis for training and new pattern recognition. Discuss these results comparing with relevant studies from the state-of-the-art. Some suggested studies can be also used for comparison.

13- Please explain Figure 7, and improve the quality of Figure 9. This last figure is too small, and it is difficult of reading.

Discussion

14- Discuss better the Figure 12, and comment the importance of this result considering patients with attention and/or memory deficits. See some related studies:

14.1) Cicerone K D 1996 Attention deficits and dual task demands after mild traumatic brain injury Brain Injury 10 79–90.

14. 2) Jeunet C, NKaoua B and Lotte F 2016 Advances in user-training for mental-imagery-based bci control: psychological and cognitive factors and their neural correlates Progress in Brain Research vol 228 (Amsterdam: Elsevier) pp 3–35

15- Add a last paragraph,g remarking the relevance, main contribution, advantages, and limitations of this research.

16- The authors selected manually 20 EEG locations for motor imagery recognition. The authors should compare their results with other studies that used all channels. Furthermore, studies that applied EEG channel selection on the same dataset should be commented. In addition, the authors should discuss about the benefit of using an automatic channel selection method in future work. Please also comment the following study: 

16.1 Papitto, Giorgio, Angela D. Friederici, and Emiliano Zaccarella. "The topographical organization of motor processing: an ALE meta-analysis on six action domains and the relevance of Broca’s region." NeuroImage 206 (2020): 116321.

Author Response

We sincerely appreciate for offering the valuable comments and suggestions, which helped us to improve the quality of the paper. The following answers addressed your comments.

As a footnote, the sequence number of the introduction section in the new script has been changed from 0 to 1.

 

Point 1: In section Introduction, the authors should comment other recent/relevant references.


Response 1: We  would  like  to  thank  for  raising  this  concern.   We  have  tried  our  best  to  address  these  issues  in  the revised  manuscript. We  add four relevant references and review those works in introduction. These four papers are  as  followed:

• Tariq, M.; M.Trivailo, P.; Simic, M. Detection of knee motor imagery by Mu ERD/ERS quantification for BCI based neurorehabilitation applications. 11th Asian Control Conference(ASCC) 2017.
• Tang, Z.; Sun, S.; Zhang, S.; Chen, Y.; Li, C.; Chen, S. A Brain-Machine Interface Based on ERD/ERS for an Upper-Limb Exoskeleton Control. Sensors 2016, 16, 2050.
• Chu, Y.; Zhao, X.; Zou, Y.; Xu, W.; Song, G.; Han, J.; Zhao, Y. Decoding multiclass motor imagery EEG from the same upper limb by combining Riemannian geometry features and partial least squares regression. Journal of Neural Engineering 2020, 17.
• Yger, F.; Berar, M.; Lotte, F. Riemannian approaches in Brain-Computer Interfaces: a review. IEEE Trans. Neural. Syst. Rehabil. Eng. 2017, 25.

 

Point 2: In addition to clinical BCI studies that used ERD/ERS for motor rehabilitation, also comment some review and research that demonstrate the potential of using ERD/ERS patterns.

Response 2: We  add two relevant references in line 23,  as  followed:

• Tariq, M.; M.Trivailo, P.; Simic, M. Detection of knee motor imagery by Mu ERD/ERS quantification for BCI based neurorehabilitation applications. 11th Asian Control Conference(ASCC) 2017.
• Tang, Z.; Sun, S.; Zhang, S.; Chen, Y.; Li, C.; Chen, S. A Brain-Machine Interface Based on ERD/ERS for an Upper-Limb Exoskeleton Control. Sensors 2016, 16, 2050.

 

Point 3: Other studies should be referenced on the lines 79-81, where the authors comment the following idea: This is a common problem with complex CNN models which result in more training time and worse real-time performance. Please discuss about the BCI latency permissible to users to feel a real closed-loop, and consequently enhance neuroplasticity.

Response 3: Thanks very much for your comments and suggestions. This work mainly focuses on MI machine learning and recognition. BCI closed-loop is our feature work which discuss in the Session 5.2, see lines during 459 to 462. The details are followed:

we plan to use a prediction tool to select features that facilitates the discovery of the most predictive features can be another direction to the improved performance [1]. That have benefit for BCI researchers to research how about interactive feedback applications and corresponding control strategies[2].

References:

• Su, R.; Hu, J.; Zou, Q.; Manavalan, B.; Wei, L. Empirical comparison and analysis of web-based cell-penetrating peptide prediction tools. Briefings in Bioinformatics 2020, 21, 408–420.
• Romero-Laiseca, M.A.; Delisle-Rodriguez, D.; Cardoso, V.; Gurve, D.; Loterio, F.; Nascimento, J.H.P.; Krishnan, S.; Frizera-Neto, A.; Bastos-Filho, T. Berlin Brain–Computer Interface—The HCI communication channel for discovery. International Journal of Human-Computer Studies 2007, 65, 460–477.

 

Point 4: In section 1. Data, specify the frequency range used to acquire the EEG data. The authors should justify why 20 EEG channels were used over the referred locations. Please cite similar motor imagery classification studies that also used these locations.

Response 1: We  add two relevant references in line 129,  as  followed.

• Lal, T.N.; Schröder, M.; Hinterberger, T.; Weston, J.; Bogdan, M.; Birbaumer, N.; Schölkopf, B. Support Vector Channel Selection in BCI. IEEE transactions on biomedical engineering 2004, 51, 1003–1010.
• Jin, J.; Miao, Y.; Daly, I.; Zuo, C.; Hu, D.; Cichocki, A. Correlation-based channel selection and regularized feature optimization for MI-based BCI. Neural Networks 2019, 118, 262–270.
• Bhattacharya, S.; Bhimraj, K.; Haddad, R.J.; Ahad, M. Optimization of EEG-Based Imaginary Motion Classification Using Majority-Voting. SoutheastCon 2017. IEEE 2017, pp. 1–5.

 

Point 5: In general, this section 2. Method should be better organized to facilitate the reader understanding, and avoid redundancy. For instance, the authors can first give an overview about the proposed system, using Figure 4 and part of the text in section 2.3. DHDANet The authors combine here present tense and past tense. I suggest to use only present tense describing the proposed recognition system. Also Figures 5 and 6, and their corresponding texts can be moved to sections 2.2 to avoid the related information disperse, and consequently facilitate the reader understanding. Notice that some variables are not defined, and sometimes other variables are defined too far of the first mention. Please define all variables, preferably close to the first mention. Also avoid to use the same variable for different meanings.


Response 5: Thanks very much for your comments and suggestions. We have moved Figure 4,5,6 to Section 3.2.  The description of the algorithm is optimized in the whole Session 3.

And some variables are specified, see line 171.

 

Point 6: Check in the subsection 2.1 Input Data the following phrases (see lines 163 and 170): "E is the number of epochs and 20 MI areas epochs"; "....of E-recorded electrodes". It is confusing. Also check the sentence on the lines 171-172 pp. 5.


Response 6: We have checked the whole paper and changed all the non-standard "epoch" to "electrode".

 

Point 7: In Eq. (1) define i and j, X1 and X2, d2. Also move Figure 4 to its first mention. Make a similar action with each figure throughout all text. The quality of Figure 4 should be improved. Also edit this figure to correct some texts.


Response 7: i and j  have been supplied the description in line 230. And the right texts in Figure 4 represent the process of calculating.

 

Point 8: Replace "T·E" by "T×E" on the line 169. Also replace "Equ" by "Eq." on the line 228. Make a similar action throughout all text.  Please define i, p and q, B, and D in subsection 2.2.1 (see lines 217-231). Make a similar action with i, j, and k on the line 253.


Response 8: We apologize for these mistakes. These are all updated, corresponding to lines 171,  230, 266 and 222 of the new menuscript.

 

Point 9: Cite studies that also support the following affirmation (see lines 282-283): "the time period for an ERD/ERS....... between 500 ms to 1 s".


Response 9: We  add two correlated references in line 285,  as  followed.

• WolpawEmail, J.R.; Boulay, C.B. Brain signals for brain–computer interfaces. Brain-computer interfaces. Springer, Berlin, Heidelberg 2009, pp.29–46.
• Pfurtscheller, G.; Neuper, C. Dynamics of sensorimotor oscillations in a motor task. Brain-Computer Interfaces. Springer, Berlin, Heidelberg 2009, pp.47–64.

 

Point 10: Various redundant texts should be removed in this section 3. Experiments and Results. It is one of the main problems found in this section. See here some redundant text that can be removed: lines 317-323, lines 332-364 in subsection 3.1, lines 375-380.


Response 10: We have improved these problems, see lines 320-324, lines 335-353, lines 355-360.

 

Point 11: Please explain Figure 7, and improve the quality of Figure 9. This last figure is too small, and it is difficult of reading.

Response 11: These two pictures have been resize optimizely.

 

Point 12: Discuss better the Figure 12, and comment the importance of this result considering patients with attention and/or memory deficits. See some related studies.


Response 12: The KU_MI dataset is not ictal subjects. In this research, We group MI illiteracy or not subjects which means the subject weacher or not do the MI experiment before.

 

Point 13: Add a last paragraph,g remarking the relevance, main contribution, advantages, and limitations of this research.


Response 13: We have added these discussions in Session 6 lines 464-466.

 

Point 14: The authors selected manually 20 EEG locations for motor imagery recognition. The authors should compare their results with other studies that used all channels. Furthermore, studies that applied EEG channel selection on the same dataset should be commented. In addition, the authors should discuss about the benefit of using an automatic channel selection method in future work.


Response 14: Yes, the automatic channel selection method is one of our feature work, please see lines 446-451.

Round 2

Reviewer 1 Report

All my suggestions and corrections have been resolved satisfactorily. Please make the following minor changes:

• Line 131: Write an introductory text under the title of Section 3 "Method".
• Line 375: "The setts of hyper parameter" -> Grammar must be checked.
• Line 398 - Write an introductory text under the title of Section 5 "Analysis and Discussion".
• Line 497:  Add title to the References section "References".

Author Response

We thank you for offering the valuable comments and suggestions, which helped us to improve the quality of the paper. We apologize for these mistakes .The following answers addressed your comments.

 

Point 1: Line 131: Write an introductory text under the title of Section 3 "Method".

Response 1: We  supplement the introduction of Section 3, see lines of 132 to 136, as followed:

In this section, we discuss the main components of our method. First, we design a dual-input preprocessing method (Sec. 3.1). Next, we exploit two custom attention mechanisms respectively for temporal and spatial feature extraction (Sec. 3.2). Finally, we discuss how we train our model from the dual-input EEG (Sec. 3.3). Figure 4 contains an overview of our method.

 

Point 2: Line 375: "The setts of hyper parameter" -> Grammar must be checked. 


Response 2: We apologize for this mistake. It’s corrected to "The suite of hyper parameters", see line 373 in the revised paper.

 

Point 3: Line 398 - Write an introductory text under the title of Section 5 "Analysis and Discussion".

Response 3: We replenish the introduction of Section 5, see lines of 397 to 399, as followed:

In this section, we prove the advantage of DHDANet by comparing the recognition effect with or without our custom attention (Sec. 5.1). Then we put forward the next step of our work based on this work (Sec. 5.2).

 

Point 4: Line 497:  Add title to the References section "References".

Response 4: We apologize for this mistake. It’s corrected see line 498.

Reviewer 3 Report

The authors attended and answered partially the reviewer concerns. Some concerns in my previous report should be clarified on the manuscript. The revised manuscript is organized better, and now provides more details. The manuscript can be still benefited after proof reading by a native English speaker. In subsection 4.2 Results, some text (see on lines 356-371) that explains the methodology used for evaluation should be moved out of this subsection. I suggest to create before a subsection, and present the methodology used for evaluation (see lines 356-371). The title of subsection 5.2. Feature Works should be checked. Some affirmations in section Conclusions are not supported by the proposed experiments, results obtained, and discussion. For instance, only a MI EEG data was used for evaluation, and four previous works were only used for comparison. Only one of these references was published after 2019. Then, it is not possible to conclude that the experimental results shown that DHDANet can outperform the best methods in the literature (see lines 470-471).  A limitation of this work is that very few studies are used for comparison. The designed experiments are not addressed to demonstrated that the proposed system can help to reduce calibration time, as a calibration step is composed of EEG data collection for training. Each subject performed a total of 100 trials per class for a period of 4 s. Then, the authors should demonstrated that a similar performance can be obtained with less trials. Reading the methodology proposed for evaluation and the results, it is not clear to understand if experiments were conducted to evaluate intra-subject and cross-subject performance. Please check in section Conclusion the phrase on the lines 488-490 "This algorithm is suitable for multi-classification tasks such as intra-subject and cross-subject EEG motor imagination, and enhances the generality of classification.

Author Response

We sincerely thank you for your careful and professional comments and suggestions, which helped us to improve the quality of the paper. The following answers addressed your comments.

 

Point 1: In subsection 4.2 Results, some text (see on lines 356-371) that explains the methodology used for evaluation should be moved out of this subsection. I suggest to create before a subsection, and present the methodology used for evaluation (see lines 356-371).

Response 1: Yes, lines 356-361 are covered in Sec. 3.1. We have deleted it. Lines 362 to 364 describe training data and testing data for this work.  Then we explain some general rules for setting hyper parameters.  It help readers to understand our experimental methods.  I hope you agree to this modification.

 

Point 2: The title of subsection 5.2. Feature Works should be checked. Some affirmations in section Conclusions are not supported by the proposed experiments, results obtained, and discussion. For instance, only a MI EEG data was used for evaluation, and four previous works were only used for comparison. Only one of these references was published after 2019. Then, it is not possible to conclude that the experimental results shown that DHDANet can outperform the best methods in the literature (see lines 470-471). 

Response 2: We would like to thank for raising this concern. We supplement two 2021 papers on MI Disk selection to illustrate the importance of our feature work on automatic-matching for BCI.

1. Yu, Z.; Li, L.; Wang, Z.; Lv, H.; Song, J. The study of cortical lateralization and motor performance evoked by external visual stimulus during continuous training.IEEE Transactions on Cognitive and Developmental Systems 2021.
2. Kim, H.S.; Ahn, M.H.; Min, B.K. Deep-Learning-Based Automatic Selection of Fewest Channels for Brain-Machine Interfaces.IEEE Transactions on Cybernetics 2021.

 

Point 3: Please check in section Conclusion the phrase on the lines 488-490 "This algorithm is suitable for multi-classification tasks such as intra-subject and cross-subject EEG motor imagination, and enhances the generality of classification.

Response3: For the sake of caution, we have removed "cross-subject". We will complement the cross-subject experiment in future work on automatic matching electrodes.