IDSRN: Interpretable Dynamic System Recurrent Network for Driver Fatigue Assessment

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Overall, the article is well-written and presents an appropriate methodology. There are some aspects that could be improved. Add citations to the statistical data presented in the introduction. Could you comment on the feasibility of performing a comparison with hybrid methods and possible metrics for evaluating the use of the algorithm for real-time applications?
In line 231, the authors mention that the work is focused on a system for civil aviation pilots. In the review of the document, I assumed the work was focused on car drivers? Is there a mistake, could you clarify this point?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The article is devoted to the development of warning systems for dangerous driving situations due to driver fatigue. This article is relevant because currently more than 1.3 million people die in road accidents worldwide, some of which are associated with micro-falling asleep at the wheel. This problem is especially relevant for truck drivers and long-distance drivers. In the interests of reducing the likelihood of drivers falling asleep and waking up, it is proposed to use electroencephalograms. At the same time, there are a number of problems related to the nonlinearity and non-stationarity of the signals received on the encephalogram. Existing methods often fail to balance recognition accuracy and computational efficiency when developing functions and models. Especially in real-time monitoring scenarios, models must provide fast and accurate identification of fatigue conditions with limited computing resources. To solve this problem, it is proposed to conduct deep learning of a neural network, namely an interpreted dynamic recurrent neural network.

The research methodology and the scheme of the presented neural network are given. I would like to note that it is necessary to show encephallograms with characteristic marks so that readers can understand how the moments of falling asleep were determined. Also, after the experimental results, it is necessary to either tell or show a diagram of how the neural network will further understand the context for people with reduced activity, or it is necessary to write about the calibration of the presented system. In graph 6, it is not clear by how many percent PGN and BP-MTN differ, since they are both greater than 90%, it is better to start with 90% on the y axis.

In general, I would like to say that the work has all the signs of a new scientific result for the prevention of road accidents, the article is well written and interesting to read, the list of references consists of relevant articles. It can be seen from Graph 7 that the presented method interprets the results of the encephalogram more accurately, which is an important achievement.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper and presented materials might be interesting to a certain set of readers, but the form and current positioning of the article must be significantly modified. 

1. Language, phrasing, and writing of this paper must be edited. It is full of minor mistakes in formation (e.g., reference list is not consistent in style),  doubtful phrasing (e.g., fatigue doesn't threaten socioeconomic stability as it was claimed on line 41), typos (e.g., "research-namely" in line 596, the "," must be used instead of "-").
2. The second issue with the writing style is the inclusion of six pages of introduction. Driver fatigue assessment is not that hard to understand and is not an unknown topic; "Introducing" it in 6 pages makes the whole paper less attractive and leaves the feeling that the authors "poured the water" in the text. Reduce the introduction to a single, maximum of two pages to identify exactly the authors' input into scientific knowledge. If, for a later discussion, references to other works and exploration of the background are needed, the additional section devoted solely to it may be added.
3. The third issue is the lack (complete absence) of description of the methods with which the authors compared their IDSRN method. It was claimed that "DT (Decision Tree), SVM (Support Vector Machine), KNN (K Nearest Neighbors), and LSTM (Long Short-Term Memory)" were used in comparison, but it is absolutely not clear 1. how those methods have been utilized, 2. Did those methods use the EEG and Eye-tracking Glasses data, or only the Eye-tracking Glasses data? Without such information, the result of the paper cannot be reproduced.
4. The fourth issue is the absence of any references to the data. If the data from Eye-tracking Glasses and EEG have been aggregated by the authors, the method of aggregation, specification, details, experiments description must be added. If the data has been aggregated by someone else, the disclosure on how to obtain this data, what the source is, and references must be added.
5. The fifth issue is the general positioning of the paper. Discussion of the usage of EEG data for the analysis of fatigue of a driver is doubtful for the same reason as the usage of MRI for the same purpose. Can such data be useful for the detection of fatigue? Sure, is it possible to make drivers put a helmet on their heads during everyday driving? Absolutely not, and we cannot expect to have a mini-MRI installed above the driver's seat. The adoption of the positioning of this paper can be redirected to bikes, ATV/UTV usage, or F1-type cars, where the usage of helmets is mandatory. Or even better, in the military case of jet planes and other aircraft, where the decrement of the pilot's conscious activity can cost millions of dollars, the price of an aircraft can easily justify an increment in the cost of the helmet for a few thousand.

At its current form, the paper leaves the reader with a very ambiguous impression; however, with significant rewriting, the addition of missing details, redirection of the application, and addressing of comments, the work can become interesting from both academic and application perspectives.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

1. Introduction – The authors included many statistical figures but did not cite any references. This should be addressed.

2. Many acronyms are mentioned without providing their full names. Please address this, or include a list of acronyms/abbreviations.

3. Dataset – This should be explained at the beginning of the Materials and Methods section. Details on the data source, as well as the training and testing datasets, should be provided.

4. The evaluation or validation metrics should be presented under a separate subheading at the end of the Methodology section.

5. Discussion – This section is missing. The authors should provide a deeper discussion of the findings, supported by relevant references.

6. Conclusion – This should clearly highlight the major findings of the study.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 5 Report

Comments and Suggestions for Authors

The authors introduce an innovative method for detecting driver fatigue that leverages electroencephalogram (EEG) signals and an Interpretable Dynamic System Recurrent Network. This combined model aims to enhance detection accuracy. Furthermore, the network architecture merges the temporal modeling capabilities of traditional Recurrent Neural Networks (RNNs) with the advantages of polynomial networks for nonlinear function approximation. This integration effectively facilitates the extraction of nonlinear dynamic features from EEG signals. Overall, the paper is well-written, and the experimental results provide promising insights for publication. My specific remarks are as follows:

- The use of polynomial expansion can result in overfitting, especially if the model picks up noise from the training data instead of the actual underlying patterns. Moreover, this approach can lead to a significant increase in the number of features, complicating both the training process and interpretation, and potentially exacerbating the "curse of dimensionality." Additionally, incorporating polynomial expansion can greatly enhance the model's complexity, making it more difficult to train and demanding more computational resources. While the architecture seeks to be interpretable, the complexity introduced by polynomial functions may hinder clarity regarding how features influence the final classification outcomes.

- As the size of the input data grows, the computational requirements of polynomial expansion may become excessive, which could hinder scalability. Furthermore, the model might face challenges in generalizing effectively to new, unseen data if it is overly tailored to the training dataset, especially in highly variable contexts like driver fatigue detection.

- It would be beneficial to include experiments conducted on edge devices to further demonstrate the model’s practical applicability.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The paper is significantly rewritten, and most of my comments have been addressed as I expected. I don't have any objections to the publication; however, a few minor corrections must be added before accepting.

1. I am not sure that the DOI for papers must be added to the reference list according to the Appl. Sci. template. But if the authors do it, the consistency is a must. Reference 4 has been listed without a DOI. Maybe in the present version of the manuscript, you can add the URL of the paper (https://arqiipubl.com/ojs/index.php/AMS_Journal/article/view/480) or its URI (http://umpir.ump.edu.my/id/eprint/40473)
   
   
2. A certain description should accompany each figure. That description would be better if it were not just a statement about the figures' titles, but some discussion on what the readers must direct their attention to. Of course, it should not be as large as a discussion in the main text, but a minimal summary, the key features, or something similar, phrased in 2-3 sentences. Such captions will make reading and referring to figures easier. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

1. Section 4 (Experimental Results) needs major revision. This section should present only the authors’ findings, but instead, it repeats the procedures. The previous review comments have not been addressed. In particular, Section 4.1 (Dataset Description) should not be included here—it belongs at the beginning of the Methodology section.
2. Section 4.6 (Discussion of Results) lacks a proper discussion. The authors have not supported their findings with any references. A thorough revision is required.
3. Additionally, the authors included more than 20 equations, but they did not cite these equation numbers in the text or provide references. All equations that were not originally developed or proposed by the authors must be properly referenced.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf