Utilizing HRV-Derived Respiration Measures for Driver Drowsiness Detection

Round 1

Reviewer 1 Report

Using Heart Rate Variability(HRV), the article investigates how to use signals obtained from wearable sensors to detect drivers' drowsiness. The state of the art uses respiratory characteristics derived from HRV signals to estimate the respiratory activity from high frequency band of HRV signals. However, the authors argued that the approach that uses dominant respiratory characteristics (DR) cannot handle situations whereby power spectrum of HRV with emphasized power at multi sub-frequency is used is unsuitable for capturing the overall respiration characteristics. They proposed a new method that uses two spectral indices, Weighted Mean (WM) and Weighted Standard Deviation (WSD) of the HF to capture the overall respiratory characteristics. In the research they found out that the proposed indices have the tendency that the frequency decreases as respiratory regularity increases in the state of drowsiness contrary to the existing solutions. The indices are then used as features to train classification algorithms to classify the state of the driver as either drowsiness or wakefulness with good performance.

·         Considering the danger posed by driver drowsiness, investigating how to detect driver's drowsiness in order to mitigate its negative impacts as it is hugely responsible for road crashes is an important research.

·         The paper is well written in terms of language and it has clearly defined the problem being addressed.

·         Authors have evaluated the proposed methods using three machine learning algorithms and comparing the results with an existing approach

The following are my comments regarding the weaknesses of this article.

·         First, even though the introduction of the article clearly defines the problem and motivates the importance of their solution, however, it is not clear why the article does not provide a section on related work to clearly situate their work in the context of the state of the art.

·         On page 3, section 2 line 88, the authors say they took 20 recordings of 6 participants through a virtual driving environment, why 20 recordings, why just 6 participants? I am afraid these small examples might have influenced the results.

·         Calculation of the proposed measure: It is difficult to read equations 1 and 2. Kindly increase the font size.

·         I am wondering why those other important sections such as methods of data collection, methods for computing HRV-derived measures, etc.  have been moved to the appendix.

Author Response

 Thank you for your kind review. We have revised the manuscript according to your valuable comments. Please see the coverletter.

Author Response File: Author Response.pdf

Reviewer 2 Report

GENERAL COMMENTS

The proposed paper deals with the identification of wake and drowsy human states during driving. The topic is interesting and also possesses reader value. However, in my opinion there some issues which should be addressed in order to make the paper suitable for publication.

In general the literature review appear quiet limited, with only 10 references reported. The introduction section needs to be supported with more literature, in order to better present the importance of the topic for a generic audience.

Secondly, the methods adopted and some assumptions and experimental choices should be clarified throughout the manuscript. In this view, in my opinion the information reported in appendix A, B and C should be moved into the methods section; otherwise, for the reader it is difficult to follow the entire experimental flow without the aspects reported at the end of the paper. Incidentally, no information is reported about the approval of the study by an ethics committee and about the information provided to the volunteers about the experimental procedures. 

The discussion section warrants supplementary information about the importance of the study outcomes and their potential practical applications, also with a more detailed comparison with the existing literature.

My detailed comments are reported below.

INTRODUCTION

- lines 39-40: please, remove the reference to the figure 1 from the intro.

- line 54: please, insert Vicente et al.

- lines 57-64: these aspects have been reported in previuos studies or are results obtained by the authors? in the former case the references are missing, in the latter case these aspects warrant to be reported as results of the study, not in the introduction section. 

- line 73: "wearable device" is quiet general, also for the introduction section. Please, specify what wearable device has been used in this study.

- figures 1, 2 and 3 refer to study results in my opinion, as I said previously. Please, consider to insert also these figures in the result section.

METHODS

- The study was approved by an ethics committee? The volunteers signed an informed consent where the experimental procedures were clearly explained? Please report this fundamental information.

- line 88: please, remove "i.e. 6 subjects", it's redundant.

- More details are needed about the virtual driving environment, due to its importance for the study design and results. Furher, a figure of the virtual driving environment should be inserted in the manuscript.

- The drowsy states of the volunteers have been assessed by the experimenter by visul inspection? This aspect it is not clear from the text, please clarify. 

- line 121: why the particular values of 40 seconds and 120 seconds have been chosen?

- lines 129-130: This sentence is unclear: what does it mean "the power spectrum is calculated with bandpass filter"?

- line 135: why did you chose the range of frequencis between 0.15 and 0.4 Hz? please, clarify.

- line 142: a difference of 0.04e-2 can hardly be considered as considerable.

- lines 145-151: this paragraph is hard to follow. Please, rephrase and clarify this procedure.

RESULTS

- line 183-184: please, rephrase the sentence.

- line 199: the greedy feedforward feature selection method deserve to be explained in more details.

- in table 2 the acronym AUC should be clarified (I suppose it stands for area under the curve).

APPENDIX A

- In my opinion the information reported in the appendix A should be inserted in the methods section, being a fundamental part of the experimental setup and not a simple additional part.

APPENDIX B

- line 265: why "or less than" is cancelled? is it only a typo?

APPENDIX C

- the limited information about prediction algorithms reported in this section does not deserve a dedicated appendix. Please, consider to move this information to the methods section.

REFERENCES

- It is not clear to what ref [6] refers: it is an article, a thesis, a project? please clarify.

- In ref [9] the journal indication is missing.

Author Response

 Thank you for your kind review. We have revised the manuscript according to your valuable comments. Please see the coverletter.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper, investigates the employment of  respiration characteristics to decide upon the wake/drowsy state of the user. Respiration characteristics are extracted from heart rate variability measure, detected from an ECG signal acquired with a Polar H7 instrument. 

The main idea of the paper is good and of great interest to the reader. The proposed application is for vehicle drivers, and the fact the driver's state of wake could be monitored with a simple, more or less low-cost, wearable monitor is of great perspective.

The introductory part and the literature review is well written

I have some concerns regarding the criteria listed in Table 1. Stability of driving: "Keeping greatly out of lane or causing big collision" - detection is performed too late! Driver behavior: "Failure to focus on driving" and "Leaning back on the chair more than 10 seconds" - how are these criteria monitored? Some references should be cited to motivate Table 1.

While the meaning of WN, defined in equation (1), is comprehensible, the definition and applicability of WSD requires more explanation. Also the values reported on the bottom of page 4, 6.60 and 7 are very close to each other and I don't consider they can be employed for classification.

On page 5, the authors refer to the scatterplots of the proposed measures and of "traditional" measures and claim: "these observations support that the proposed measures can well capture the characteristics of respiration associated with drowsiness" . In my opinion, Figures 4 and 5 show the opposite of this claim. First, In Figure 4 which shows the  scatterplot of the proposed measures, the wake state is overlapping the drowsy state making their discrimination impossible (see subjects 1 ,2 and 5), or only partially possible (see subjects 3 and 4). In comparison, Figure 5 illustrating the scatterplots of the traditional measures allows a better discrimination (see subjects 3, 4, and 5).

On the top of page 7, please check the meaning of the word model. I believe the authors wanted to write classifier... Also, the referred figure is 6, not 7  .

On page 8, the authors claim again that "the use of proposed measures can clearly improve the prediction performance in all the three types of models". Figure 6 illustrates the contradictory. As observed for the SVM and KNN classifier, HRV aids both the proposed and the dominant respiration features. Without HRV, the proposed and dominant respiration features yield comparable  results. The proposed method only works well for Random forest, which however exhibits the weakest results among the three classification methods.  

As such, the statements asserted in the conclusion are contradicted.

The appendices help to the understanding of the paper, I would suggest including them into the text for better readability.

Author Response

 Thank you for your kind review. We have revised the manuscript according to your valuable comments. Please see the coverletter.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors addressed all my previous concerns. I have no further comments. Thank you

Author Response

 Again, thanks for your kind reviews.

Reviewer 3 Report

The authors have addressed my remarks properly. Ambiguity is considerably reduces and the concepts are better explained.

However, I am not convinced by the results. I stick to my opinion that the proposed measures perform worse than the traditional measures available in literature. The newly added figure 6 confirms my belief. As illustrated, the scatter plot in Figure 6 A allows for a worse discrimination compared to Figure 6 B. 

If possible, the authors should provide additional results.

Author Response

 Thank you for your kind review. We have revised the manuscript according to your comment. Please see the coverletter.

Author Response File: Author Response.pdf