A Novel Method for ECG Signal Compression and Reconstruction: Down-Sampling Operation and Signal-Referenced Network

Round 1

Reviewer 1 Report

The comments are attached. 

Comments for author File: Comments.pdf

Author Response

Detailed Response to Reviewers

We gratefully thank the editor and all referees for their time spend making their constructive remarks and useful suggestions, which has significantly raised the quality of the manuscript and has enable us to improve the manuscript. Each suggested revision and comment, brought forward by the referees was accurately incorporated and considered. We revised our manuscript, and we would be very grateful if the referees could give a review on the revised manuscript.

 

 

1.The author should further elaborate on the choice of down-sampling rate and the trade-off between compression and signal quality. It would also be useful to compare the performance of the proposed method with other existing ECG signal compression and reconstruction techniques. Additionally, a discussion on the limitations and potential applications of the method could provide more context for the results presented.

 

-Reply：

Thanks for your comments. The trade-off between compression and signal quality is highly related to the application scenarios of the wearable ECG monitoring system. We comprehensively evaluated the proposed method to have a detailed reference for readers and their projects instead of a specific application.

Having said that, we give a practical application of our project. In our project, the wearable ECG monitoring system is used to mainly detect Atrial fibrillation (AF), and the detection of Premature atrial contraction (PAC) and Premature ventricular contraction (PVC) is of secondary importance. However, in our project, the working duration of the wearable device is short, and its battery cannot be changed. Therefore, the objective of using the proposed method is to extend the working duration of the wearable device and maintain the detection capacity of the system for the main disease. When the wearable device has a high-level battery, the sampling frequency is 500Hz. If the battery becomes lower, the sampling frequency decreases to 250Hz. In this state, the reconstructed signals may not have high quality enough to provide information on PAV and PVC but still can be used to detect AF from its RR intervals, therefore the main disease detection still can be guaranteed. If the battery becomes lower, the sampling frequency decreases to 125Hz or less. Finally, the working duration of the wearable device is extended and the main disease detection is also guaranteed. Additionally, we also add this application in the discussion part.

 

2.The English writing level of the manuscript should be improved.

-Reply：

Thanks for your comments. We have optimized the English writing by using a third-party service.

 

3.Keywords should be in alphabet order and consistent in terms of capital or lowercase;

-Reply：

Thanks for your comments. We have corrected them in the revised manuscript.

 

4.Introduction: this section should be improved to show the advantages/disadvantages of references. In addition, it is highly recommended and strongly suggested that the authors add the following reference to support the context of Wearable Exoskeleton and their functionality.

https://doi.org/10.1109/TNSRE.2021.3136088

 

-Reply：

Thanks for your comments. This paper has a deep insight into the Wearable Exoskeleton field. We would be very honored if our work can promote the progress of this field. We have added this reference to our manuscript.

 

5. Section 2.1. The authors should provide more detail on the signal-referenced network, such as the architecture and training process. Additionally, the choice of the referencing signal and its impact on the performance of the signal-referenced network could be further discussed.

 

-Reply：

Thank you for the comment. In this section, we mainly introduce the context of the proposed method. The purpose is to show where the down-sampling operation can be applied, where the signal can be reconstructed using the signal-referenced network, and what the level of decreasing the data size is. And then, we introduce the down-sampling operation and the signal-referenced network in detail and separately in sections 2.2 and 2.3. Therefore, we provide more detail about the training process with the pseudo code in section 2.3. Additionally, we add a discussion about the importance of the referencing signal. And, the pseudo-code shows below:

Pseudo Code of Training Process

1.Initialize model, GPU, Optimizer

2.Load data for DS_train, DS_validate, DS_test.

3.For i=1 in range of 100 epochs:

Train model with DS_train, calculate loss, and backward propagation.

Validate model with DS_validate, obtain MSE.

If MSE is smaller than last one:

Save current parameters of the model as the best ones.

6.Load the best parameters of the model.

7.Test model with DS_test and get the final result.

 

6. Section 2.2. The authors are recommendedto provide more detail on the choice of the compression rate and its impact on the quality of the compressed signals. Additionally, the method used to choose the compression rate and the trade-off between compression and signal quality could be further discussed. Overall, the section provides a good foundation for understanding the down-sampling operation and the method used to compress ECG signals.

-Reply：

Thanks for your comment. According to the compressing equation, the larger the compressing rate, the more sparse the signal, and the more difficult the reconstruction. It is very testing of the reconstruction performance of the proposed network. For heart diseases with the oblivious feature, for example, the ventricular premature beat with an early RR interval and a wide QRS complex, the device can run with a higher compressing rate. For the ischemic cardiovascular disease, its diagnosis always requires high quality of ECG signals, the device must run with a lower compressing rate. Therefore, the trade-off trade-off between compression and signal quality must be considered according to the application.

 

7. Section 2.3. The authors should provide more detail on the evaluation metrics used to assess the performance of the network and the results of the evaluation. Additionally, a discussion on the limitations and potential areas of improvement of the network would provide more context for the results presented.

-Reply：

Thank you for your comments.  We have noticed that there were many evaluation metrics to evaluate the effectiveness of the reconstructed signal, for example, the width of the QRS complex, the duration of P and T waves, and the amplitude of the R peak. In our opinion, the definition of the effectiveness of the reconstructed signal is different in different situations. When the reconstructed signal is sent to a classification algorithm, we can consider the reconstructed signal effective if the classification algorithm gives a correct result even though the duration and amplitude have large error. Therefore, the F1 score is a straightforward evaluation metric for the reconstruction effectiveness evaluation. And, the RMSE can well quantify the effectiveness. Overall, these two metrics are complementary and can give a straightforward evaluation of the reconstruction effectiveness. Moreover, we conclude with a discussion of the limitations of the proposed method and an outlook for its future applications at the end of Chapter 4.

 

8. Section 3. Please address the following fro this section: - Provide more detail and explanation about the experimental design. For example, why were ten experimental schemes chosen and what was the reasoning behind dividing them into four groups?

 

-Reply：

Thank you for your comments. These four groups can demonstrate the necessity of the design of the proposed method. The first group demonstrates that which lead is the best one for reconstructing signals. The second group demonstrates the necessity of the reconstructed signals for the classification algorithm. The third group demonstrates the necessity of using a referencing signal. The forth group demonstrates the necessity that keeping the original sampling frequency of the referencing signal.

 

9.Clearly explain how the datasets were selected and how they were split into training, validation, and testing sets. Additionally, it would be helpful to provide more information about the China Physiological Signal Challenge 2018 database and how the records from it were used.

-Reply：

Thank you for your comments. It is very important to choose and separate the database. And, It determines whether the experimental results are effective or not. We employed the classification algorithm to select signals from the database available for this work. This algorithm is the second prize in the CPSC 2018 and can be downloaded from the official website. This essential operation avoids lacking authenticity of the result due to artificial subjectivity. There is also a principle that the selected data must keep the ECG classification algorithm’s F1 score of 100% on all classes. This principle excludes uncertainty made floating on the result of the classification and makes the reasons for the floating of results directly pointing to the effectiveness of the reconstructed signals. Moreover, keeping the independence of each separated dataset, including the training set, validating set, and testing set, is also very important. It determines whether the results are true or not. We utilized the CPSC 2018 for some reasons. The first is that the sampling frequency and resolution rate are the same as our device, and the data is more abundant than ours. The second is that due to the restriction of our private data, the public data can be used to compare in a better way. In this work, we only used the “Training set” on the official website. This dataset can be obtained with the link “http://2018.icbeb.org/Challenge.html”. We have added supplementary information about the database in our revised manuscript.

 

9. Provide a more in-depth explanation of the evaluation metrics used. For example, what is the significance of Root Mean Square Error (RMSE) and F1-score in evaluating the proposed method?

 

-Reply：

Thank you for your comments. We have noticed that there were many evaluation metrics to evaluate the effectiveness of the reconstructed signal, for example, the width of the QRS complex, the duration of P and T waves, and the amplitude of the R peak. In our opinion, these metrics are too scattered to evaluate the effectiveness of the proposed method. For example, although the duration of the QRS complex has a small error and the amplitude of other waveforms of a heartbeats has a large error, the classification algorithm can output a correct result. We cannot regard the reconstructed signal as ineffective due to the poor performance of reconstructing other waveforms. Therefore, for this situation, the F1 score is a very straightforward metric for evaluate the effectiveness of the reconstructed signals. To quantify the effectiveness, we employed the metric RMSE. Overall, the F1 score validates the effectiveness of the reconstructed signal, and the metric RMSE can quantify the effectiveness of the reconstructed signal.

 

10. Include results and findings from the experiments, as well as a discussion of the implications of these results. This would make the section more comprehensive and provide valuable information to the reader.

-Reply：

Thank you for your comments. We are sorry for the incorrect title of this section. In this section, we mainly introduce the experiment design and the evaluation metrics. We thank you for your suggestions. In Section 4, we have added more discussions about the possible application and the findings.

 

11. Consider including visual aids, such as graphs or charts, to help illustrate the results and findings.

-Reply：

Thanks for your comment. To demonstrate the authenticity of our work, we have added measurement results related to the power consumption of our device in our revised manuscript. And, we also add some graphs of our wearable device and the power consumption measurement in our revised manuscript.

 

12. Clarity of Figures: The figures used to present the results could benefit from clearer labeling and more concise captions to help the reader better understand what is being shown.

-Reply：

Thanks for your comment. We have tried to regenerate Figure 9 with high resolution. The new figure has added in the revised manuscript.

 

 

13. Discussion: The discussion of the results could be more thorough and provide more context. For example, it would be helpful to compare the results of the proposed method to other existing methods in the field and provide more insight into why the proposed method performed well or not as well in certain cases.

-Reply：

Thanks for your comment. To our best knowledge, our method is the first work which uses a deep learning network to reconstruct the signal and uses the down-sampling operation to compress signals. In essence, the compress progress of our method is quite different to other methods. For other methods, the compress progress needs algorithms to realize the compression and the compressed content is future matrices not signals. At this point, it is not suitable to compare our method with other compression methods. Even so, we will make access of our data and codes opening, and we welcome more researchers to propose their methods and compare them with our methods. Moreover, due to the black box feature of deep learning, some cases with poor performance cannot be explained. But, we are now trying our best to develop an interpretable network base on current work to explain such cases.

 

14.Limitations: The limitations section could be expanded to provide a more detailed discussion of the limitations of the proposed method and how they could be addressed in future work.

 

-Reply：

Thanks for your comment. To our best knowledge, this work is the first attempt to use a deep learning network to reconstruct the signal and uses the down-sampling operation to compress signals. Although the performance is in line with our expectations, it still has many limitations. The first limitation is the fixed sampling frequency. We plan to investigate the relationship between the sampling frequency and the ECG signal quality. When the ECG signal has good quality, the sampling frequency decreases to lower. And, when the ECG signal has poor quality, the sampling frequency keeps high to retain the detail of the signal. Moreover, we also plan to propose a more flexible structure for the network in our next work. The second limitation is that the quality of the reconstructed signal is high relying on the train signals. We will expand the types of training data and improve the quality of training data. Meanwhile, we will validate the possibility that the network can remove noises when reconstructing signals. We also add the related supplement to the limitation section in the revised manuscript.

 

Author Response File: Author Response.pdf

Reviewer 2 Report

Electronics - A Novel Method for ECG Signal Compression and Reconstruction: Down-sampling Operation and Signal- referenced Network

 

Specific and Thorough Comments to the Authors

 

Recommendation: Accept after minor revision

 

The authors present a method for ECG signal compression and reconstruction aimed at reducing power consumption in wireless communication. While the idea is promising, I have concerns regarding several aspects, as outlined below:

 

1.     The authors should validate the feasibility of practical measurement applications, rather than relying solely on the existing database.

 

2.     The authors should assess the impact of various down-sampling scales on power consumption. If the authors can explicitly demonstrate the amount of power saved by the proposed method during wireless communication, it would significantly enhance the contribution to wearable applications.

 

3.     Similar to Question 2, in addition to Bluetooth, the authors can employ the proposed method to quantify the power savings in Wi-Fi mode.

 

4.     Regarding your F1-score results, do they include ECG data from patients with heart disease? If so, the authors should discuss the F1-score results specifically for different subject groups.

 

 

5.     Can the proposed method be applied to arrhythmia signals? Is it capable of reconstructing most of the signal after down-sampling? Are there any limitations to the current method?

Author Response

Detailed Response to Reviewers

We gratefully thank the editor and all referees for their time spend making their constructive remarks and useful suggestions, which has significantly raised the quality of the manuscript and has enable us to improve the manuscript. Each suggested revision and comment, brought forward by the referees was accurately incorporated and considered. We revised our manuscript, and we would be very grateful if the referees could give a review on the revised manuscript.

 

 

1.The authors should validate the feasibility of practical measurement applications, rather than relying solely on the existing database.

 

-Reply：

Thanks for your comments. We utilized the CPSC 2018 for some reasons. The first is that the sampling frequency is the same of our device, and the data is more abundant than ours. The second is that due to the restrict of our private data, the public data can be used to compare in a better way. In this work, we only used the “Training set” in the official website. We will make the access of our data and codes opening, and we welcome more researchers to propose their methods and compare with our methods.

 

2. The authors should assess the impact of various down-sampling scales on power consumption. If the authors can explicitly demonstrate the amount of power saved by the proposed method during wireless communication, it would significantly enhance the contribution to wearable applications.

-Reply：

Thanks for your comments. The proposed method has been applied to our system. We add the measurement to our revised manuscript. The result shows below. In this table, the current is averaged and measured within 1 hour. The basic current refers that the BLE being in the broadcasting state and no data transmitting. In this table the current decreases with the decrease in the sampling frequency.

Table V The Current Measurement of the Proposed Method

Sampling Frequency (Hz)	25	50	125	250	500
Current (mA)	6.07	6.23	6.46	7.13	7.83
Basic Current (mA)	5.85

 

 

3. Similar to Question 2, in addition to Bluetooth, the authors can employ the proposed method to quantify the power savings in Wi-Fi mode.

-Reply：

Thanks for your comments. In our device, the chip is nRF52832, it does not support WiFi, therefore, we can only give the measurement result of the BLE. We have considered your comment seriously, and we plan to use a wireless chip that supports WiFi and BLE and give the measurement result in our next work.

 

4. Regarding your F1-score results, do they include ECG data from patients with heart disease? If so, the authors should discuss the F1-score results specifically for different subject groups.

-Reply：

Thanks for your comments. Our database used in this work includes various arrhythmia. The result F1 score is the average of all arrhythmia. Of curse, we have noticed that the sampling frequency affects the F1 score of different arrhythmia and explained the reason according to Table III. However, the factors affecting the result include many aspects, including the method of down-sampling signals, the compressing rate, the performance of the reconstruction network, the classification algorithm, and the imbalance database. Facing these factors, there are many places needed o to be explored. For all this, we are writing a new paper to analyze the impact of different factors on the results, and we will give possible ways to solve this problem.

 

5. Can the proposed method be applied to arrhythmia signals? Is it capable of reconstructing most of the signal after down-sampling? Are there any limitations to the current method?

-Reply：

Thank you for the comment. The result of our method shows that our method can reconstruct some types of arrhythmia, including Atrial fibrillation (AF), First-degree atrioventricular block (I-AVB), Left bundle branch block (LBBB), Right bundle branch block (RBBB), Premature atrial contraction (PAC), Premature ventricular contraction (PVC), ST-segment depression (STD), ST-segment elevated (STE). However, such a method is now in its infancy, some arrhythmia has not been validated. We will validate more arrhythmia in our next work by using more comprehensive databases. Although the performance is in line with our expectations, it still has many limitations. The first limitation is the fixed sampling frequency. We plan to investigate the relationship between the sampling frequency and the ECG signal quality. When the ECG signal has good quality, the sampling frequency decreases to lower. And, when the ECG signal has poor quality, the sampling frequency keeps high to retain the detail of the signal. Moreover, we also plan to propose a more flexible structure for the network in our next work. The second limitation is that the quality of the reconstructed signal is high relying on the train signals. We will expand the types of training data and improve the quality of training data. Meanwhile, we will validate the possibility that the network can remove noises when reconstructing signals. We also add the related supplement to the limitation section in the revised manuscript.

 

 

Author Response File: Author Response.pdf

Reviewer 3 Report

Xiaojun et al have produced and effective method to compress and reconstruct the ECG signal and transmit and reconstruct the signal effectively. The development of ambulatory medical equipment are critical and challenging The manuscript is well presented and interesting. I suggest following clarifications before being considered for publication.

1. The experimental instruments is missing. Authors should add the detail regarding the set up.

2. Remove the link in figure 1 and instead add schematic presenting details of the experimental work.

3. One major problem with ECG is the noise and body movement. This paper should show it is possible to use this mechanism while person is moving otherwise a wire can be used. The ECG sometimes is very hard to recover during hard movement so at least authors should mention the speed and physical activity possible.

4. The future direction regarding wireless implants should be discussed.

5. Authors should briefly discuss the power saving as a result of BLE utilization related to a life span of a commercial portable battery compared with other classes of Bluetooth that use high power. Also low power may result in a safe usage of the technology avoiding/less localize heat generation during usage of BLE.

6. regarding the algorithm by stacking or low frequency conversion, would there be any latency?

7. are the sampling frequency optimization performed when the patient is stationary? Would that change if patient is moving?

 

Author Response

Detailed Response to Reviewers

We gratefully thank the editor and all referees for their time spend making their constructive remarks and useful suggestions, which has significantly raised the quality of the manuscript and has enable us to improve the manuscript. Each suggested revision and comment, brought forward by the referees was accurately incorporated and considered. We revised our manuscript, and we would be very grateful if the referees could give a review on the revised manuscript.

 

 

1. The experimental instruments is missing. Authors should add the detail regarding the set up.

 

-Reply：

Thanks for your comments. In our work, the wearable device was developed by our team. Its sampling frequency is 500 Hz and can work for 3 days continuously. It has only 17g and can acquire 3-channel ECG signals. We have measured the current of our devices with different down-sampling frequencies. The results show in Table V. The current measurement instrument is KEITHLEY DMM7510, and the sampling frequency for the current measurement is 1KHz. The Figure 10, it shows our wearable device.

Table V The Current Measurement of the Proposed Method

Sampling Frequency (Hz)	25	50	125	250	500
Current (mA)	6.07	6.23	6.46	7.13	7.83
Basic Current (mA)	5.85

 

 

 

 

Figure 10 A diagram of measuring current

 

 

 

 

2. Remove the link in figure 1 and instead add schematic presenting details of the experimental work.

-Reply：

Thanks for your comments. We have edited this Figure in our revised manuscript.

Figure 1. A structure diagram of our wearable ECG monitoring system. (a) The signals’ transmission path. (b) The schematic of the server.

 

 

3. One major problem with ECG is the noise and body movement. This paper should show it is possible to use this mechanism while person is moving otherwise a wire can be used.The ECG sometimes is very hard to recover during hard movement so at least authors should mention the speed and physical activity possible.

 

-Reply：

Thanks for your comments. To our best knowledge, this work is the first attempt to use a deep learning network to reconstruct the signal and uses the down-sampling operation to compress signals. Although the performance is in line with our expectations, it still has many problems, for example, the noise problem. Our wearable monitoring system had been used in our collaborating hospital. In this hospital, the patients are long in the tooth, and they walk slowly. Therefore, the ECG signals have mild noise. However, it is a serious problem that the signal is very hard to be reconstructed during strenuous movement. But, our method has an advantage. It is the deep learning network. For the reconstructed signal, the quality is relied on the quality of the training data. Therefore, there is a huge space to improve the signal quality by optimizing the network and improving the quality of training signals.

 

4. The future direction regardingwireless implants should be discussed.

-Reply：

Thanks for your comments. Wireless implants have the potential to revolutionize the field of healthcare by providing real-time monitoring and control of various bodily functions. In this field, some problems still needed to be further solved. One of these problems is miniaturization. For this problem, reducing the need for battery capacity is a possible solution. Our method has been applied to the practical project and extends the duration of the wearable device. Therefore, our method may be a power technology that can Promote the development of wireless implants.

 

5. Authors should briefly discuss the power saving as a result of BLE utilization related to a life span of a commercial portable battery compared with other classes of Bluetooth that use high power. Also low power may result in a safe usage of the technology avoiding/less localize heat generation during usage of BLE.

 

-Reply：

Thank you for the comment. Our device used the chip nRF52832. It only supports BLE, not classic Bluetooth. The full name of BLE is Bluetooth Low Energy, which is developed for power-limited applications and devices. The difference between BLE and other Bluetooth is that some power-costing functions are removed from classic Bluetooth and it has a lower communicating speed. The basic communicating speed can only meet the wearable device. Nevertheless, the consumption of BLE still considerable when the data size is large. In the result in Table IV, the current decreases from 7.83mA to 6.07mA when the sampling frequency decreases to 25 Hz, the devices can work with another several hours. It is very practical no mater for the commercial or some emergencies.

 

6. regarding the algorithm by stacking or low frequency conversion, would there be any latency?

 

-Reply：

Thank you for the comment. The down-sampling operation is just to skip some sampling points. It do not involve any algorithms, has no need to compute the algorithm, therefore, only the data size becomes smaller and there is no latency.

 

7. are the sampling frequency optimization performed when the patient is stationary? Would that change if patient is moving?

 

-Reply：

Thank you for the comment. In this work, the sampling frequency is set according to the diseases of the patient, therefore, it is fixed. But we were enlightened by your questions. If the sampling frequency can changed automatically according to the movement, the diseases, or other conditions, the reconstructed signals may be better and the power consumption may be lower. Therefore, we are now trying our best to realize it. Thank you.

 

 

 

 

 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Dear Editor,

I have thoroughly reviewed the paper and I must say that it has improved significantly since the last time I reviewed it. The revisions that th e authors have made have greatly enhanced the clarity and cohesiveness of the paper. It is evident that they have put a lot of effort into addressing the feedback that was provided.

After carefully considering the content of the paper, I highly recommend it for publication. The paper provides valuable insights and presents a well-supported argument. The research is sound, the methodology is appropriate, and the results are clearly presented.

Overall, I am very impressed with the quality of your paper, and I believe that it will make a valuable contribution to the field.