Comparative Analysis of Deep Learning Models for Optimal EEG-Based Real-Time Servo Motor Control

Round 1

Reviewer 1 Report

This study introduces a deep learning approach combining CNN and LSTM for Motor Imagery using EEG signal recordings. The research is compelling and yields promising results. However, several aspects require clarification and improvement:

1. Section Organization: Section 2.3 should be divided into distinct sections such as "2.3 Data Preprocessing", "2.4 Feature Extraction", and "2.5 Models" for better readability and structure.

2. Signal Segmentation and Feature Extraction: There is a need for detailed clarification on how the EEG signals were segmented and the exact number and nature of features extracted. Additionally, specifying the Python libraries used for these processes would be beneficial.

3. Results Section: The writing in the results section should be strengthened. In some cases, information currently spread across multiple single-sentence paragraphs can be consolidated into a single, more cohesive paragraph.

4. Dataset Availability: It should be clarified whether the datasets used in the study are publicly available, providing transparency and enabling reproducibility.

5. Model Availability: Similarly, it is important to state if the models developed in the study are accessible to the public.

By addressing these points, the paper can improve its clarity, readability, and reproducibility, making it more valuable to the research community.

The English in this paper is of good quality, showcasing great grammatical accuracy and a diverse vocabulary.

Author Response

We would like to extend our heartfelt thanks to the reviewers for their thorough and insightful feedback on our manuscript. Your constructive comments and suggestions have been invaluable in refining and enhancing the quality of our work. We have carefully addressed each point raised and made significant revisions to improve the clarity, structure, and scientific rigor of our study.

Reviewer 1

1.  Section Organization: Section 2.3 should be divided into distinct sections such as "2.3 Data Preprocessing", "2.4 Feature Extraction", and "2.5 Models" for better readability and structure.
• Response: Section 2.3 has been reorganized into distinct sections for better readability and structure: "2.3 Data Preprocessing," "2.4 Feature Extraction," etc. Also, section 1.1 "Related work" has been added.
2. Signal Segmentation and Feature Extraction: There is a need for detailed clarification on how the EEG signals were segmented and the exact number and nature of features extracted. Additionally, specifying the Python libraries used for these processes would be beneficial
• Response: Detailed information on how the EEG signals were segmented and the exact number and nature of features extracted has been provided. We have also specified the Python libraries used for these processes.
3. Results Section: The writing in the results section should be strengthened. In some cases, information currently spread across multiple single-sentence paragraphs can be consolidated into a single, more cohesive paragraph.
• Response: The results section has been strengthened by consolidating information that was previously spread across multiple single-sentence paragraphs into cohesive paragraphs.
4. Dataset Availability: It should be clarified whether the datasets used in the study are publicly available, providing transparency and enabling reproducibility.
• Response: It has been clarified that the datasets used in the study are publicly available. The Data Availability Statement includes a link to the datasets.
5. Model Availability: Similarly, it is important to state if the models developed in the study are accessible to the public
• Response: The Model Availability Statement has been added, explaining that the models are not publicly available due to proprietary considerations. However, interested researchers can contact the corresponding author for potential collaborations or access under
  specific conditions.

Reviewer 2 Report

While the manuscript covers the necessary aspects of feature extraction from preprocessed EEG signals, incorporating the suggested elaborations and clarifications would significantly improve the clarity and depth of the discussion. The text is well-structured and covers critical steps in the preprocessing phase.

Reported below are some suggestions for improvement to enhance clarity and comprehensiveness:

• The discussion would be complete by presenting an image of the entire frequency spectra before and after filtering, as it visually demonstrates the effectiveness of the preprocessing steps in isolating the relevant EEG components while eliminating noise. This visualization would significantly enhance the reader's understanding.
• The description of the preprocessing steps is thorough. However, the sequence could be better organized to ensure a logical flow.
• The choice of the fourth-order Butterworth bandpass filter and its frequency range is well justified. It would be helpful to include a brief rationale for selecting the Butterworth filter over other filters, emphasizing its advantages for this specific application.
• The text mentions using a specialized artifact removal algorithm, the Discrete Wavelet Transform (DWT). It would be beneficial to briefly compare DWT and other artifact removal methods and highlight why DWT was chosen.
• The manuscript mentions both time-domain and frequency-domain features, a comprehensive approach. However, it would benefit from a more precise differentiation between the two and the rationale behind choosing specific features for motor imagery tasks.
• It would be helpful to include more detail on how the power spectral density ratio was computed and its specific relevance to motor imagery tasks.
• Hjorth parameters are well-known for providing insights into EEG signal characteristics. The manuscript could be strengthened by explaining how these parameters relate to motor imagery and what specific insights they can provide about cognitive processes during such tasks.
• The Petrosian fractal dimension is a valuable measure of signal complexity. Including examples or references to previous studies where this measure has been used successfully in motor imagery would provide more context and validation.
• The discussion of the Frobenius norm could be expanded to explain its application in more detail, including how it differentiates cognitive conditions and its specific role in the analysis.
• Describing spectral power and its relevance to motor imagery tasks is straightforward. However, more detail on how spectral power was computed and any preprocessing steps applied before its calculation would be beneficial.
• Wavelet transform coefficients are mentioned for capturing transient features. It would be helpful to specify which wavelet functions were used and why they were chosen for this particular analysis.
• The manuscript mentions frequency band ratios (e.g., alpha/beta, theta/beta) and their relevance to different cognitive states. Providing more detail on calculating these ratios and their specific significance in the context of motor imagery tasks would strengthen the discussion.
• The manuscript mentions correlating frequency-domain features with servo motor movements' speed, direction, and precision. This is an important aspect that could be elaborated further. Including specific examples or case studies where these correlations were observed would enhance the reader's understanding of the research's practical applications.

The manuscript provides a detailed and comprehensive description of the methodology for training deep learning models using k-fold cross-validation on EEG data.

Reported below are some suggestions for improvement to enhance clarity and comprehensiveness:

• The rationale for choosing k=5 is explained, focusing on balancing computational demands and performance estimation variance. However, citing relevant literature or empirical evidence supporting this choice would strengthen the argument.
• The importance of stratified sampling in maintaining class distribution balance is well-explained. It would be beneficial to briefly mention how this specific dataset's class distribution influenced the decision to use stratification.
• The Adam optimizer and sparse categorical cross-entropy loss function are justified for multi-class classification on EEG data. It might be helpful to briefly mention how these choices align with the characteristics of EEG data.
• GridSearchCV for hyperparameter tuning is appropriately mentioned. A brief example of the hyperparameter grid explored could illustrate the range of parameters considered and the rationale behind their selection.

With these few suggestions, providing visual aids and more specific examples would significantly enhance the comprehensiveness and impact of the manuscript, making it more comprehensive.

Author Response

We would like to extend our heartfelt thanks to the reviewers for their thorough and insightful feedback on our manuscript. Your constructive comments and suggestions have been invaluable in refining and enhancing the quality of our work. We have carefully addressed each point raised and made significant revisions to improve the clarity, structure, and scientific rigor of our study.
Reviewer 2

1. The discussion would be complete by presenting an image of the entire frequency spectra before and after filtering, as it visually demonstrates the effectiveness of the preprocessing steps in isolating the relevant EEG components while eliminating noise. This visualization would significantly enhance the reader's understanding
• Response: An image of the entire frequency spectra before and after filtering has been included to visually demonstrate the effectiveness of the preprocessing steps.

2. The description of the preprocessing steps is thorough. However, the sequence could be better organized to ensure a logical flow.

• Response: The sequence of preprocessing steps has been organized has been reorganized into distinct sections for better readability and structure.

3. The choice of the fourth-order Butterworth bandpass filter and its frequency range is well justified. It would be helpful to include a brief rationale for selecting the Butterworth filter over other filters, emphasizing its advantages for this specific application.

• Response: A brief rationale for selecting the Butterworth filter over other filters has been included, emphasizing its advantages for this specific application.

4. The text mentions using a specialized artifact removal algorithm, the Discrete Wavelet Transform (DWT). It would be beneficial to briefly compare DWT and other artifact removal methods and highlight why DWT was chosen.

• Response: A comparison between DWT and other artifact removal methods has been included, highlighting why DWT was chosen.
1. The manuscript mentions both time-domain and frequency-domain features, a comprehensive approach. However, it would benefit from a more precise differentiation between the two and the rationale behind choosing specific features for motor imagery tasks.
• Response: A precise differentiation between time-domain and frequency-domain features has been provided, along with the rationale behind choosing specific features for motor imagery tasks.

6. It would be helpful to include more detail on how the power spectral density ratio was computed and its specific relevance to motor imagery tasks.

• • Response: More detail on how the power spectral density ratio was computed and its relevance to motor imagery tasks has been included.

7. Hjorth parameters are well-known for providing insights into EEG signal characteristics. The manuscript could be strengthened by explaining how these parameters relate to motor imagery and what specific insights they can provide about cognitive processes during such tasks.

• • Response: An explanation of how Hjorth parameters relate to motor imagery and the specific insights they provide about cognitive processes during such tasks has been added.

8. The Petrosian fractal dimension is a valuable measure of signal complexity. Including examples or references to previous studies where this measure has been used successfully in motor imagery would provide more context and validation.

• Response: Examples and references to previous studies where the Petrosian fractal dimension has been used successfully in motor imagery have been included to provide more context and validation.

9. The discussion of the Frobenius norm could be expanded to explain its application in more detail, including how it differentiates cognitive conditions and its specific role in the analysis.

• Response: The discussion of the Frobenius norm has been expanded to explain its application in more detail, including how it differentiates cognitive conditions.

10. Describing spectral power and its relevance to motor imagery tasks is straightforward. However, more detail on how spectral power was computed and any preprocessing steps applied before its calculation would be beneficial.

• Response: Details on how spectral power was computed and any preprocessing steps applied before its calculation have been provided.

11. Wavelet transform coefficients are mentioned for capturing transient features. It would be helpful to specify which wavelet functions were used and why they were chosen for this particular analysis.

• Response: The wavelet functions used in the analysis have been specified, and the rationale behind their choice has been explained.

12. The manuscript mentions frequency band ratios (e.g., alpha/beta, theta/beta) and their relevance to different cognitive states. Providing more detail on calculating these ratios and their specific significance in the context of motor imagery tasks would strengthen the discussion.

• Response: More detail on calculating frequency band ratios and their specific significance in the context of motor imagery tasks has been included.

13. The manuscript mentions correlating frequency-domain features with servo motor movements' speed, direction, and precision. This is an important aspect that could be elaborated further. Including specific examples or case studies where these correlations were observed would enhance the reader's understanding of the research's practical applications.

• Response: The discussion on correlating frequency-domain features has been elaborated further, Ιncluding case studies where these correlations were observed.

Additional Points

1. The rationale for choosing k=5 is explained, focusing on balancing computational demands and performance estimation variance. However, citing relevant literature or empirical evidence supporting this choice would strengthen the argument.

• Response: The rationale for choosing k=5 in k-fold cross-validation has been explained, with citations to relevant literature supporting this choice.

2. The importance of stratified sampling in maintaining class distribution balance is well-explained. It would be beneficial to briefly mention how this specific dataset's class distribution influenced the decision to use stratification.

• Response: The importance of stratified sampling in maintaining class distribution balance has been briefly mentioned, noting how the specific dataset's class distribution influenced this decision.

3. The Adam optimizer and sparse categorical cross-entropy loss function are justified for multi-class classification on EEG data. It might be helpful to briefly mention how these choices align with the characteristics of EEG data.

• Response: The choice of the Adam optimizer and sparse categorical cross-entropy loss function has been justified in alignment with the characteristics of EEG data.

4. GridSearchCV for hyperparameter tuning is appropriately mentioned. A brief example of the hyperparameter grid explored could illustrate the range of parameters considered and the rationale behind their selection.

• Response: A brief example of the hyperparameter grid explored in GridSearchCV has been included to illustrate the range of parameters considered and the rationale behind their selection.