An Innovative Deep Learning Algorithm for Drowsiness Detection from EEG Signal

Round 1

Reviewer 1 Report

The paper proposes to detect the drowsiness stage in EEG signal using auto-encoder based method. The authors show the method, the results and conclude the paper. The topic is interesting,however,the paper needs much more work before accepting it. Some comments that give support for my decision are below:

1- The ref. [5] came after the [17]. It should be in order;

2- At line 47, there is a period at the wrong place;

3- I guess that the following phrase is out of context: "Yann LeCun and Yoshua Bengio were the first to apply in 1998 the Convolutional Neural Networks (CNNs) in computer  vision [13]."

4- Don'tlines146 and 148 represent equations? If so, they should number it. The equations are in terrible qualities? Can't you use the Equation to write the Equations? There is no standard about that.

5- From line 190 to 199 there are several small paragraphs. Is that true?

6- I wasn't able to see the labels in Figure 3. The authors should improve the quality of the figures. Does Fig 3(c) represent DCT spectral for Fig 3(a) or 3(b). They should include it in both the text and caption.

7- The author could add Fig 3 (d) for another state;

8- Fig 4 needs several improvements. They also should put at the same figure the signals for both states, for each encoder;

9- The authors should include information about time consumption for both, training and validation.

10- The authors used just a few samples for training the system. How did they prevent the over-fitting procedure? Could they go further in that?

11- All figures need improvement in their qualities. They should provide bigger letters.

12- The authors have developed much research related to the theme. Why there is no comparison with other methods?

13- The authors should improve the introduction. Based on the quick search, there is a lot of related work. Some of them involve deep learning based approach. Some of them are below:

https://0-doi-org.brum.beds.ac.uk/10.1016/j.aap.2017.11.038

https://0-doi-org.brum.beds.ac.uk/10.1016/j.image.2016.05.018

RODNEY PETRUS BALANDONG ,RANA FAYYAZ AHMAD ,MOHAMAD NAUFAL MOHAMAD SAAD, AND AAMIR SAEED MALIK, A Review on EEG-Based Automatic Sleepiness Detection Systems for Driver ,IEEE access. 

Prediction of driver's drowsy and alert states from EEG signals with deep learning ,Conference: Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), 2015 IEEE 6th International Workshop, December 2015 DOI: 10.1109/CAMSAP.2015.7383844 .

Driver drowsiness detection using ANN image processing, T.Vesselenyiet al2017 IOP Conf. Ser.: Mater. Sci. Eng. 252 012097

Tingxi Wen  ;Zhongnan Zhang  et al. Deep Convolution Neural Network andAutoencoders-Based Unsupervised Feature Learning of EEG Signals, IEEE access,vol6, 25399 - 25410, 2018.

Author Response

Referee 1

REFEREE 1

The paper proposes to detect the drowsiness stage in EEG signal using auto-encoder based method. The authors show the method, the results and conclude the paper. The topic is interesting,however,the paper needs much more work before accepting it. Some comments that give support for my decision are below:

-The ref. [5] came after the [17]. It should be in order;

Author’s answer:

Thank you for the comment. We corrected the references number and re-ordered them correctly.

2- At line 47, there is a period at the wrong place;

Author’s answer:

Thank you for the comment. We corrected the period in the line 47.

3- I guess that the following phrase is out of context: "Yann LeCun and Yoshua Bengio were the first to apply in 1998 the Convolutional Neural Networks (CNNs) in computer vision [13]."

Author’s answer:

Thank you for the comment. We rephrased the period to better explain the concept we addressed.

4- Don't lines146 and 148 represent equations? If so, they should number it. The equations are in terrible qualities? Can't you use the Equation to write the Equations? There is no standard about that.

Author’s answer:

Thank you for your comment. We rewrote all the equations in editable format and numbered them accordingly.

5- From line 190 to 199 there are several small paragraphs. Is that true?

Author’s answer:

Thank you for your comment. We corrected the periods in line 190-199 merging them in a single paragraph.

6- I wasn't able to see the labels in Figure 3. The authors should improve the quality of the figures. Does Fig 3(c) represent DCT spectral for Fig 3(a) or 3(b). They should include it in both the text and caption.

Author’s answer:

Thank you for your comment. We improved the quality of figure 3 according to your suggestion.

7- The author could add Fig 3 (d) for another state;

Author’s answer:

Thank you for your comment. We added 3d according to your suggestions.

8- Fig 4 needs several improvements. They also should put at the same figure the signals for both states, for each encoder;

Author’s answer:

Thank you for your comment. We improved the figure 4 according to your suggestions by adding additional “features” according to suggestions of the referee n.2

9- The authors should include information about time consumption for both, training and validation.

Author’s answer:

Thank you for your comment. We added a new paragraph entitled “2.4 Algorithm Testing and Validation framework” in Material Method Section explaining how training and validation were carried out.

10- The authors used just a few samples for training the system. How did they prevent the over-fitting procedure? Could they go further in that?

Author’s answer:

Thank you for your comment. To prevent the overfitting issues we have implemented an ad hoc regularized empirical risk function as well as we have used Kullback-Leibler divergence function (see the end of the paragraph 2.3 - b) AutoEncoder -The Autoencoder System Block – raws 217-222).

11- All figures need improvement in their qualities. They should provide bigger letters.

Author’s answer:

Thank you for your comment. We improved the quality of all figures according to your suggestions.

12- The authors have developed much research related to the theme. Why there is no comparison with other methods?

Author’s answer:

Thank you for your comment. We have added two comparison tables (Table 1 and Table 2) and discussion on that at the end of the Results and Discussion Sections.

13- The authors should improve the introduction. Based on the quick search, there is a lot of related work. Some of them involve deep learning based approach.

Some of them are below:

https://0-doi-org.brum.beds.ac.uk/10.1016/j.aap.2017.11.038

https://0-doi-org.brum.beds.ac.uk/10.1016/j.image.2016.05.018

RODNEY PETRUS BALANDONG,RANA FAYYAZ AHMAD ,MOHAMAD NAUFAL MOHAMAD SAAD, AND AAMIR SAEED MALIK, A Review on EEG-Based Automatic Sleepiness Detection Systems for Driver,IEEE access.

Prediction of driver's drowsy and alert states from EEG signals with deep learning,Conference: Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), 2015 IEEE 6th International Workshop, December 2015 DOI: 10.1109/CAMSAP.2015.7383844 .

Driver drowsiness detection using ANN image processing, T.Vesselenyiet al2017 IOP Conf. Ser.: Mater. Sci. Eng. 252 012097

Tingxi Wen  ;Zhongnan Zhang  et al. Deep Convolution Neural Network andAutoencoders-Based Unsupervised Feature Learning of EEG Signals, IEEE access,vol6, 25399 - 25410, 2018.

Author’s answer:

Thank you for your comment. We improved the introduction according to your comment and suggested references by adding a text reported in the raws 75-98.

Reviewer 2 Report

This manuscript reports research on DEEP LEARNING ALGORITHM FOR DROWSINESS DETECTION FROM EEG SIGNAL. The authors need to address some of the major concerns before the paper is considered for publication.

Specific comments to each section:

Introduction section need to be further improved with more explanation and literature review. This study uses EEG for data analysis and interpretation, but authors have not mentioned recent literature related to EEG fatigue classification and artifacts removal procedure related to the study. Techniques such as ICA/BSSS are widely used for the same. Please cite the following papers related to artifact removal.

Kasakawa, Shinya, et al. "Approaches of Phase Lag Index to EEG Signals in Alzheimer’s Disease from Complex Network Analysis." Innovation in Medicine and Healthcare 2015. Springer International Publishing, 2016. 459-468.

Chai, Rifai, et al. "Improving EEG-Based Driver Fatigue Classification Using Sparse-Deep Belief Networks." Frontiers in neuroscience 11 (2017).

Liu, Guotao, et al. "Complexity Analysis of Electroencephalogram Dynamics in Patients with Parkinson’s Disease." Parkinson’s Disease 2017 (2017).

Al-Ani, Ahmed, Irena Koprinska, and Ganesh Naik. "Dynamically Identifying Relevant EEG Channels by Utilizing Their Classification Behaviour." Expert Systems with Applications (2017).

Chai, Rifai, et al. "Driver fatigue classification with independent component by entropy rate bound minimization analysis in an EEG-based system." IEEE journal of biomedical and health informatics 21.3 (2017): 715-724.

Jia, Huibin, Huayun Li, and Dongchuan Yu. "The relationship between ERP components and EEG spatial complexity in a visual Go/Nogo task." Journal of neurophysiology 117.1 (2017): 275-283.

Feature extraction methods need to be explained in detail including the artifact removal procedure. Please cite the following papers related to artifact removal.

S. Bhardwaj et al, "Online and automated reliable system design to remove blink and muscle artefact in EEG," 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Milan, 2015, pp. 6784-6787.

Kwon, Younghee, et al. "Efficient learning of image super-resolution and compression artifact removal with semi-local Gaussian processes." IEEE transactions on pattern analysis and machine intelligence 37.9 (2015): 1792-1805.

P. N. Jadhav et al, "Automated detection and correction of eye blink and muscular artefacts in EEG signal for analysis of Autism Spectrum Disorder," 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Chicago, IL, 2014, pp. 1881-1884.

The rationale behind the choice of different features and classifier combinations need to be clearly explained in the manuscript.

The analysis proposed in this research need to be compared with the other state of the art methods including other entropy methods.

Please improve the quality of the figures and also include ROC curve, specificity and sensitivity analysis.

Discussion and conclusion section need to be further improved. 

Author Response

REFEREE 2

This manuscript reports research on DEEP LEARNING ALGORITHM FOR DROWSINESS DETECTION FROM EEG SIGNAL. Thes authors need to address some of the major concerns before the paper is considered for publication.

Specific comments to each section:

1. Introduction section need to be further improved with more explanation and literature review. This study uses EEG for data analysis and interpretation, but authors have not mentioned recent literature related to EEG fatigue classification and artifacts removal procedure related to the study. Techniques such as ICA/BSSS are widely used for the same. Please cite the following papers related to artifact removal.

Kasakawa, Shinya, et al. "Approaches of Phase Lag Index to EEG Signals in Alzheimer’s Disease from Complex Network Analysis." Innovation in Medicine and Healthcare 2015. Springer International Publishing, 2016. 459-468.

Chai, Rifai, et al. "Improving EEG-Based Driver Fatigue Classification Using Sparse-Deep Belief Networks." Frontiers in neuroscience 11 (2017).

Liu, Guotao, et al. "Complexity Analysis of Electroencephalogram Dynamics in Patients with Parkinson’s Disease." Parkinson’s Disease 2017 (2017).

Al-Ani, Ahmed, Irena Koprinska, and Ganesh Naik. "Dynamically Identifying Relevant EEG Channels by Utilizing Their Classification Behaviour." Expert Systems with Applications (2017).

Chai, Rifai, et al. "Driver fatigue classification with independent component by entropy rate bound minimization analysis in an EEG-based system." IEEE journal of biomedical and health informatics 21.3 (2017): 715-724.

Jia, Huibin, Huayun Li, and Dongchuan Yu. "The relationship between ERP components and EEG spatial complexity in a visual Go/Nogo task." Journal of neurophysiology 117.1 (2017): 275-283.

Feature extraction methods need to be explained in detail including the artifact removal procedure. Please cite the following papers related to artifact removal.

S. Bhardwaj et al, "Online and automated reliable system design to remove blink and muscle artefact in EEG," 2015 37th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Milan, 2015, pp. 6784-6787.

Kwon, Younghee, et al. "Efficient learning of image super-resolution and compression artifact removal with semi-local Gaussian processes." IEEE transactions on pattern analysis and machine intelligence 37.9 (2015): 1792-1805.

P. N. Jadhav et al, "Automated detection and correction of eye blink and muscular artefacts in EEG signal for analysis of Autism Spectrum Disorder," 2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Chicago, IL, 2014, pp. 1881-1884.

Author’s answer:

Thank you for your comment. We improved the introduction section according to your comment and suggested references by adding a text reported in the raws 38-41.

2. The rationale behind the choice of different features and classifier combinations need to be clearly explained in the manuscript.

Author’s answer:

Thank you for your comment. We explained the rationale we have used for the choice of different features and classifier combinations by adding a text in the section 2.3 Algorithm Description - a) DCT - Discrete Cosine Transform block raws 154-157

3. The analysis proposed in this research need to be compared with the other state of the art methods including other entropy methods.

Author’s answer:

Thank you for your comment. We have added two comparison tables (Table 1 and Table 2) and discussion on that at the end of the Results and Discussion Sections.

4.Please improve the quality of the figures and also include ROC curve, specificity and sensitivity analysis.

Author’s answer:

Thank you for your comment.

We improved the quality of all figures according to your suggestions.

Concerning the ROC curve, specificity and sensitivity analysis, we highlighted this aspect by adding a specific text after the figure 6 (raws 312-315).

5.Discussion and conclusion section need to be further improved. 

Author’s answer:

According to your suggestion, we improved the discussion section (see text yellow highlighted).

Round 2

Reviewer 1 Report

The authors have properly improved the manuscript.

Reviewer 2 Report

The authors have addressed all my comments satisfactorily and the paper can be considered for publication.