Involuntary Breathing Movement Pattern Recognition and Classification via Force-Based Sensors

Round 1

Reviewer 1 Report

This study investigates to develop pattern recognition and classification scheme to detect IBMs through force sensors. Overall, this manuscript was well organized and need of study is important to address clinical needs. As the study mentions in further studies, limitations such as lung volumes and more details must be addressed for significance or any correlations of output results.

Major comment:

1.     Include reasons for choice of tests for each analysis (Kolmogorov–Smirnov test, Levene’s test and others). Preferable cite papers from literature.

Here are the minor comments from my side:

1.     Introduce abbreviations in introduction again. or add all of them at end of paper for ease

2.     Include test type in Fig 2 legend.

3.     Figure 3. Color bar selection just use hot-color bar as current looks more blacker levels compare to other.

4.     Legends are misleading. I would suggest A square markers for clusters. In addition, could you reorganize them P1-P8 in orderly fashion. Also talk about “P4” about cluster 2 and its NM and IBM.

Author Response

Major comment:

1. Include reasons for the choice of tests for each analysis (Kolmogorov–Smirnov test, Levene’s test, and others). Preferably cite papers from the literature.

Res: We have cited the literature. We have also mentioned in the methodology that the data is non-normal distributed. Therefore, we used nonparametric tests such as the KS test and another nonparametric test. We understand the reviewer’s suggestion, and we have added citations and statements regarding the choice of test.

Minor Comments

1. Introduce abbreviations in the introduction again. or add all of them at end of the paper for ease

Res: Thanks, we have made the changes as asked. All of the relevant abbreviations are mentioned in the introduction.

2. Include test type in Fig 2 legend.

Res: Thanks, we have added the test and relative citations in the stats analysis subsection and have also added the test type for the p-value.

3. Figure 3. Color bar selection just uses the hot-color bar as current looks blacker levels compare to others.

Res: Dear reviewer, we have used a hot color bar but we have a lot of data points for the NM in comparison to the IBM phase, Thus, the overall graphs give a darker illusion. We wanted to show the darker area as it represents the NM phase.

4. Legends are misleading. I would suggest A square markers for clusters. In addition, could you reorganize them P1-P8 in an orderly fashion? Also, talk about “P4” about cluster 2 and its NM and IBM.

Res: Dear Reviewer, we have made changes as asked. We have organized the panels in such a way so that they are well sequenced.

One of the participant’s NM phases got clustered into IBM due to noise or fluctuations in the data. Thus, our clustering is not 100% accurate but almost 99% accurate, especially when splitting NM and IBM phases as ground truth. We thank the reviewer for raising this point. I have now mentioned this in our text.

Reviewer 2 Report

This paper presents a pattern recognition for identifying the sub-phase of involuntary breathing movement (IBM) phase during breath-holding (BH) using eight subjects.  The k-means is to category the IBM phase, and then SVM, NB,  DT and K-NN to identify the labeled data.  The paper is interests.  However, there are many problems.

(1) what is the contribution of the paper?  please clarify it.

(2)Based on the Section 2.1, the raw data should be collected from force plates.  However, what is the force plates? how to collect the raw breathing data?  what is the sensor, it is unclear.  

Based on the code, authors uses the EMG package to deal with the data, but there has no any information of EMG in the paper.

(3)Section 2.7 introduces the supervised learning.  It is unclear how to get the ground truth on the sub-phase?  is it from k-means output?

(4)Fig.2 shows only two phases.  Then why did paper cluster five-six optimal clusters?

(5) What is the. x and  y-axis of Fig.3, it is unclear  why author provide these figures?

(6)   as k-means is unsupervised, how did authors understand the six-cluster is the optimal in Fig.4? 

(7)Figure 5 shows that cluster should be 5, but fig.4 shows 6 ?

(8)Fig.8 shows BH from two subjects. It is mentioned that x-axis is normalised time.  As x-axis is a time how authors normalise time? In general,  y-axis should be normalised? why authors did not do it?

(9) It is mentioned that "The code is for processed data and uses test data to validate model accuracy. https://github.com/rajatsingh91/Classification-Recognition-Convulsion". However, the code is just a function.  There has no any data example.  In addition, the codes were confused as it is used the EMG package.   Authors should clarify it. 

Author Response

Reviewer 2

1. What is the contribution of the paper? please clarify it.

RES: Thank you for your comment, in the third and fourth paragraphs we have stated what is the research gap and how our study fills that gap (contribution).

“The current literature does not present methods to recognize subphases within the IBM phase. Therefore, methods emphasizing detecting patterns or sub-phases within the IBM phase are of importance in clinical science and should be developed. Hence, the goal of our study is to design such a method. We refer to the IBM phase as several events or patterns of IBMs during the SP. In this study, we present a novel scheme (shown in figure 1) that can recognize different sub-phases within the IBM phase during BH, and classify them accordingly.”

2. Based on Section 2.1, the raw data should be collected from force plates.  However, what are the force plates? how do collect the raw breathing data?  what is the sensor, it is unclear. 

Res: We thank the reviewer for his comment on this matter, we have added the statement regarding force plates containing load cells that measure force associated with a movement.

3. Section 2.7 introduces supervised learning.  It is unclear how to get the ground truth on the sub-phase.  is it from the k-means output?

Res. We do not have the ground truth for subphases but we have ground truth for the NM phase and IBM phases. The cluster analysis can cluster these phases nicely. The classifiers can classify these phases easily. Moreover, the idea of doing unsupervised learning is to assume about subphases as unsupervised learning relies on identifying patterns without having prior knowledge about the data. Whereas, in supervised learning, we already know the input to the model. Therefore, we made assumptions about the subphases from k means and fed that information into the classifier to classify them. 

4. Based on the code, the authors use the EMG package to deal with the data, but there has no information on EMG in the paper.

Res: We used the EMG biosignal package for signal processing. In our case, we used the EMG package for enveloping. The enveloping does the same job on the force signal as it does on the EMG signal because the algorithms for signal processing do not change much with different packages. The only thing that will change is the parameters of the enveloping algorithm, such as the time window for root mean square or moving average. We made sure that those parameters are set up based on the frequency of the signal which we have mentioned in the methods. We are adding a note in the paper that states this rationale.

5. 2 shows only two phases.  Then why did the paper cluster five-six optimal clusters?

Res: Figure 2 shows the ground truth or two main phases which are statistically significant. It is important to validate it through statistics as we used visual inspection to categorize these two phases. We clustered these two phases and also subphases within the IBM phase.

It is crucial to identify the optimal number of clusters before defining the number of clusters in k means. This is a general method in unsupervised learning to identify prior to defining cluster numbers clusters. We can identify it based on the total within the sum of squares and we have already mentioned that within the text. I am also attaching a tutorial about our methodology to identify an optimal number of clusters. Kindly find attached a tutorial link “https://www.r-bloggers.com/2017/02/finding-optimal-number-of-clusters/”

5. What is the. x and y-axis of Fig.3, it is unclear why the author provides these figures.

Res: Dear reviewer, On the top right panel, I have shown two coordinates x-axis is acceleration, and the y-axis is a jerk. We have now stated this statement in the caption of the figure. These figures provide a relationship between acceleration and jerk. This relationship is important to understand the phasic nature of IBM.

6. 2 shows only two phases.  Then why did the paper cluster five-six optimal clusters?

Res: There are two phases of movement and there are several events or subphases of movement within these two phases. We are clustering all these events in 6 clusters. Additionally, the results do not change much if you cluster them in either 5 clusters or six clusters.

7. Figure 5 shows that the cluster should be 5, but fig.4 shows 6.

Res: Dear Reviewer, figure 5 shows how increasing the number of clusters affects the performance of a certain classifier (Decision tree) as the k means clusters output are used as labels for the decision tree. The increase in the number of clusters from 5-6 does not impact the results that much as mentioned earlier. However, increasing it further reduces the performance of the decision tree. Therefore, figure 4 shows different results than figure 6.

8. 8 shows BH from two subjects. It is mentioned that the x-axis is normalized time.  As the x-axis is a time how do authors normalize time? In general, the y-axis should be normalized. why authors did not do it?

Res: The only reason we normalized time for the x-axis is to visually represent and compare data between two participants. We used interpolation to normalize time, and our goal was to show different phases and how IBM fluctuations change over time in two different participants. Normalization of force will not affect our claims but it will not show the peculiarity between two participants’ fluctuating IBMs.

9. It is mentioned that "The code is for processed data and uses test data to validate model accuracy. https://github.com/rajatsingh91/Classification-Recognition-Convulsion". However, the code is just a function.  There has no data example.  In addition, the codes were confused as it is used in the EMG package.   The authors should clarify it. 

Res. Dear Reviewer. Thank you for your comment but the data cannot be shared based on our protocol on public platforms. Kindly contact the corresponding author for it. Moreover, we have used many packages to design the code in R. Therefore, you have to look at the documents for those functions that we have used. We have already presented our logic in the form of a block diagram, and we have used functions to design our complete scheme and functions are generally a part of a code.