Enhancing Classification Accuracy with Integrated Contextual Gate Network: Deep Learning Approach for Functional Near-Infrared Spectroscopy Brain–Computer Interface Application
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsIn this study, an integrated contextual gate network (ICGN) algorithm is proposed to enhance the classification accuracy. The results show that ICGN can be efficiently used for the classification of two and three-class problems in fNIRS-based BCI applications. However, there are some problems should be considered.
The research depth of this paper is limited, (1) only two classes of fNIRS data, (2) lack of further discussion in the proposed algorithm, and (3) be not compared with some latest related research.
The information of ethical approval document should be added in Line 106.
There is no substantial difference in the content expressed between Figure 6 and Figure 7, since that the accuracy plus the loss equals 100%.
How about the relationship between the parameters of proposed algorithm and the performance of model?
Besides accuracy, some other evaluation indexes such as training time can be demonstrated in this paper.
Author Response
Manuscript No.: Sensors-2913602
Manuscript Title: Enhancing Classification Accuracy with Integrated Contextual Gate
Network: Deep Learning Approach for fNIRS BCI ApplicationAuthors: Jamila Akhter, Noman Naseer, Hammad Nazeer, Haroon Khan and Peyman
Mirtaheri 1. Summary
The authors would like to thank the Editor and two anonymous reviewers for their valuable time in reviewing and critiquing our manuscript. The manuscript has been revised upon the reviewers’ comments. The authors’ point-by-point answers to the comments are provided below.
- Point-by-point response to
Response to Reviewer 1
Comment #1: The research depth of this paper is limited, (1) only two classes of fNIRS data, (2) lack of further discussion in the proposed algorithm, and (3) not compared with some latest related research.
Response: Additional discussion has been added to address the limitation of the study in terms of class count, algorithm discussion, and comparison with the latest research. Specifically, detailed results of validation using a three-class problem have been added in Section 4 (Pages 12-17) of the revised manuscript. These enhancements enrich the paper's content and contribute to its overall depth and relevance.
Comment #2: The information of ethical approval document should be added in Line 106.
Response: The ethical approval document number has been added (Page 2, Line 89), ensuring transparency and compliance with ethical standards. Also, an ethical approval document has already been submitted along with the manuscript.
Comment #3: There is no substantial difference in the content expressed between Figure 6 and Figure 7, since the accuracy plus the loss equals 100%.
Response: The reviewer is right. In a two-class classification scenario, where the classes are well-separated, we might expect to see similar trends in accuracy and loss graphs, as both metrics directly relate to the model's ability to differentiate between the two classes. Therefore, the loss graph for two-class problem has been removed from the revised manuscript. However, in a three-class classification problem, the complexity increases, potentially resulting in higher losses due to the added difficulty of distinguishing between multiple classes. Consequently, the similarity in accuracy and loss graphs between binary and multiclass problems may differ significantly. It underscores the importance of considering the nature of the classification task when interpreting model performance metrics. The accuracy and loss graphs of model performance for three classes of problems are added in the revised manuscript in Section 4 (Pages 12-17).
Comment #4: How about the relationship between the parameters of the proposed algorithm and the performance of the model?
Response: Discussion on the relationship between the parameters of the proposed algorithm and model performance has been added (Page 17, lines 368-383). This analysis might enhance the reader's understanding of how algorithm parameters impact classification accuracy and provide insights into potential optimization strategies.
Comment #5: Besides accuracy, some other evaluation indexes such as training time can be demonstrated in this paper.
Response: The training and testing computational time comparison between LSTM, Bi-LSTM, and the proposed ICGN algorithm has been included, along with plots, to provide additional evaluation indexes (Page 11 for 2-class and Page 17 for 3-class). These additions can offer a comprehensive assessment of the proposed algorithm's performance across multiple metrics. Thank you.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors1. In the abstract, dataset details shall be presented before the discussion of the result.
2. The abstract does not present the challenges focused on in the paper in addition to manual feature extraction using deep learning presented in many literatures. Therefore, it can not be defined as a challenge or contribution.
3. Sentences is not clear: In functional near-infrared spectroscopy-based 12 BCI (fNIRS-BCI) to enhance classification performance deep learning (DL) algorithms are used to substitute manual feature extraction.
4. The author should discuss how fMRI’s temporal resolution is handled.
5. It is confusing that if the model is trained and tested for two-class classification then how validation is performed using a three-class dataset?
6. The test results for the two classes are presented well but validation results are not discussed in detail. Detailed experiment results are required.
7. The main objective is to replace the manual feature extraction process using a deep network for better classification. But Author did not present a comparison and analysis of improvement between them.
8. A comparative analysis of the literature is expected.
Author Response
Manuscript No.: Sensors-2913602
Manuscript Title: Enhancing Classification Accuracy with Integrated Contextual Gate
Network: Deep Learning Approach for fNIRS BCI ApplicationAuthors: Jamila Akhter, Noman Naseer, Hammad Nazeer, Haroon Khan and Peyman
Mirtaheri 1. Summary
The authors would like to thank the Editor and two anonymous reviewers for their valuable time in reviewing and critiquing our manuscript. The manuscript has been revised upon the reviewers’ comments. The authors’ point-by-point answers to the comments are provided below.
- Point-by-point response to
Response to Reviewer 2
Comment #1: In the abstract, dataset details shall be presented before the discussion of the result.
Response 1: Thank you for your insightful comment. Dataset details have been presented before the discussion of the results in the abstract of the revised manuscript (Page 1, lines 14-16). This change would help the readers gain a clear understanding of the dataset used in the study before the results are discussed.
Comment #2: The abstract does not present the challenges focused on in the paper in addition to manual feature extraction using deep learning presented in many literatures. Therefore, it cannot be defined as a challenge or contribution.
Response: The main contribution of the current paper is in the development of the ICGN algorithm and the demonstration of its effectiveness in feature extraction which leads to improved classification performance. This has been clarified in the revised manuscript. It has been emphasized that the novelty lies in the ICGN algorithm's ability to outperform existing methods in feature extraction, thereby, enhancing classification accuracy.
Comment #3: Sentences are not clear: In functional near-infrared spectroscopy-based BCI (fNIRS-BCI) to enhance classification performance, deep learning (DL) algorithms are used to substitute manual feature extraction.
Response: The sentence regarding the use of deep learning algorithms in fNIRS-BCI has been rewritten for better clarity (Page 1, lines 12-13). The revised sentence provides a clear explanation of how deep learning algorithms perform better for feature extraction as compared to manual feature extraction for fNIRS-BCI, which leads to better classification performance.
Comment #4: The author should discuss how fMRI’s temporal resolution is handled.
Response: Clarification has been provided that the current study uses fNIRS, not fMRI, and detailed explanation of fNIRS has been included.
Comment #5: It is confusing that if the model is trained and tested for two-class classification then how validation is performed using a three-class dataset?
Response: First the model undergoes training and testing for two-class problem. After that, to verify the effectiveness of the proposed algorithm, an open-access dataset containing three-class problem has been utilized. The model is then trained again on the open-access three-class dataset and its accuracies have been determined. This validates the effectiveness of our algorithm for two-class as well as three-class problems. A separate section (Pages 12-17) with detailed results of three-class problem has been added in the revised manuscript.
Comment #6: The test results for the two classes are presented well but validation results are not discussed in detail. Detailed experiment results are required.
Response: The manuscript now incorporates comprehensive validation results (Pages 12-17) for open-access three-class datasets. These additional details offer a thorough insight into the model's performance across diverse datasets, thereby, augmenting the overall presentation of the paper.
Comment #7: The main objective is to replace the manual feature extraction process using a deep network for better classification. But the author did not present a comparison and analysis of improvement between them.
Response: The core objective of this study is to introduce the Integrated Contextual Gate Network (ICGN), a novel deep learning algorithm. ICGN aims to enhance classification accuracy while minimizing computational expenses. Compared to conventional RNN algorithms like LSTM and Bi-LSTM, ICGN exhibits superior performance owing to its advanced pattern identification and feature extraction methods from input datasets. Its innovation lies in utilizing contextual information to extract more insightful features, resulting in enhanced performance across various classification tasks.
Comment #8: A comparative analysis of the literature is expected.
Response: The discussion section now incorporates a comparative analysis of relevant literature (Page 18, lines 349-362 and 372-374). This analysis provides valuable context for the research, illustrating the novelty and significance of the proposed approach within the existing body of knowledge.
Thank you.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper has been revised according to the suggestion. It can be considered to accept.
Author Response
Thank you for accepting our paper. We appreciate your time and expertise in reviewing our work.
