Efficient Real-Time R and QRS Detection Method Using a Pair of Derivative Filters and Max Filter for Portable ECG Device

Round 1

Reviewer 1 Report

General overview:

This manuscript proposes a method for real-time QRS detection in ECG signals acquired by portable equipment. The manuscript is, overall, very well written and organised, and both the description of the method and its evaluation are very thorough. I just leave a few comments that the authors should consider.

Strengths: (1) very good motivation; (2) complete description of the method; (3) thorough evaluation;

Weaknesses: (1) the novelties of the algorithm and differences with derivative-based state-of-the-art methods should be clearer in the introduction; and (2) the manuscript could be shorter.

Comments:

The authors should have made the novelties of the proposed method clearer in the introduction. It would be beneficial to have a discussion of the differences between the proposed method and other state-of-the-art methods based on derivatives;

Section 2.1 offers an explanation on arrhythmias. I believe it is not really needed to establish the motivation behind this work, and it is clearly in the wrong place as it does not describe materials nor methods.

The evaluation with the MIT-BIH Arrhythmia database is very complete and thorough, as well as the test with ECG patch recordings. However, I wonder if the authors have pondered on testing the algorithm on off-the-person ECG databases (like UofTDB or CYBHi). Efficient QRS detection is very important in off-the-person ECG-based biometric recognition, and this method could be a very good alternative to existing methods.

In Table 2, it is unclear how the authors assessed the real-time feasibility of the state-of-the-art methods. Is it based on time limits? On computational complexity limits? The authors should explain this.

 Figure 12 is divided over two pages. It should be fixed.

Author Response

The authors should have made the novelties of the proposed method clearer in the introduction. It would be beneficial to have a discussion of the differences between the proposed method and other state-of-the-art methods based on derivatives;

Sol) According to your advice, the following explanation is added in the introduction.
(Line 88-100) Recently, derivative-based QRS detection methods are still being studied due to their low complexity [29, 30]. Zhang et al. [29] have presented a pulse triggered R peak detection algorithm for ECG signals based on the second derivative. This method uses a QRS morphology for R peak detection and is computationally inexpensive. However, in this method, the detection of Q wave, a trough in ECG signal, should be preceded before R peak detection. Therefore, if Q wave is contaminated by noise or there is no trough point of Q wave, the R point detection performance is significantly lowered. Rivas et al. [30] proposed a QRS detector using derivation and adaptive thresholding. In this method, the accuracy of initial maximum peak detected as the R peak greatly affects the overall performance and has a disadvantage of being sensitive to noise. The derivative-based method is in the spotlight again with development of wearable ECG devices, however, it is sensitive to noise. In addition, the existing methods do not have a criterion for determining whether detected R peak is true R peak or noise due to absence of noise detection function, which may cause errors in the RR interval calculation and further HRV analysis.
29. Zhang, X.; Lian, Y. A 300-mV 220-nW event-driven ADC with real-time QRS detection for wearable ECG sensors. IEEE Transactions on Biomedical Circuits and Systems 2014, 8, 834–843.
30. Gutiérrez-Rivas, R.; García, J.J.; Marnane, W.P.; Hernández, Á. Novel real-time low-complexity QRS complex detector based on adaptive thresholding. IEEE Sensors Journal 2015, 15, 6036–6043.

Section 2.1 offers an explanation on arrhythmias. I believe it is not really needed to establish the motivation behind this work, and it is clearly in the wrong place as it does not describe materials nor methods.

Sol) Yes, right. The section 2.1 explaining the arrhythmia was moved to the introduction.

The evaluation with the MIT-BIH Arrhythmia database is very complete and thorough, as well as the test with ECG patch recordings. However, I wonder if the authors have pondered on testing the algorithm on off-the-person ECG databases (like UofTDB or CYBHi). Efficient QRS detection is very important in off-the-person ECG-based biometric recognition, and this method could be a very good alternative to existing methods.

Sol) We performed additional experiment for CYBHi database as the following.
(Line 291-299) CYBHi DB (1000 samples/sec) [31,32], off-the-person data, is also used for verifying the R point detection performance of the proposed method. The signals of the DB contain more noises compared to the MIT-BIH DB. The sliding window length of 3000 samples and the step size of 1500 samples are used for the experiment. ECG signals of the CYBHi DB used in the experiments were filtered by a low-pass filter with a 40 Hz cut-off frequency. Figure 15 shows the R point detection result for Record 20110719-RMAF-CI-8B, 20110718-ARS-CI-8B, and 20110718-ARS-CI-85 section in the CYBHi DB. While 20110719-RMAF-CI-8B section contains bigger T waves, 20110718-ARS-CI-8B section and 20110718-ARS-CI-85 section contain DC wandering and bigger P waves. Nevertheless, it can be seen that the proposed method detects the R point relatively accurately.
31. Silva, H.P.; Lourenço, A.; Fred, A.; Raposo, N.; Aires-de-Sousa, M. Check your biosignals here: a new dataset for off-the-person ECG biometrics. Comput Methods Programs Biomed. 2014, 113, 503–514.
32. Check your biosignals here initiative (CYBHi) dataset for off-the-person electrocardiography (ECG) biometrics. Available online: https://zenodo.org/record/2381823#.XYn59kYzabh (accessed on 24 Sept. 2019).

In Table 2, it is unclear how the authors assessed the real-time feasibility of the state-of-the-art methods. Is it based on time limits? On computational complexity limits? The authors should explain this.

Sol) That is based on time limit. The following sentence is added in the explanation of Table. 2.
(Line 313-314) The criterion of real-time feasibility means that the average calculation time for the 1-second length ECG signal of Record 111 is less than 10 ms.

Figure 12 is divided over two pages. It should be fixed.
Sol) OK, Figure 12 was divided into (a)~(b) and (c)~(e).

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper presents an alternative real-time algorithm for the detection of R and QRS information based on the application of derivative filters to ECG signals. The proposed method improves the QRS intervals and the classification of R peaks. Furthermore, a noise detection method has been designed to remove specific patterns of noises that could affect the quality of ECG signals. The performance of the proposed method was evaluated with respect to other studies in literature and time processing was also estimated.

The paper is interesting as well as the application in real-time ECG monitoring.

I have some comments and questions for authors that can improve the overall quality of the paper, the readability and the understanding for the readers. 

In my opinion, abstract contains too many details on methods that are not needed in this section. I think that authors should modify it and “balance” its structure (background, objective of the study, methods, results and conclusions) for greater impact on the reader. L43-L52: some additional references on available technologies could be inserted to improve the state of the art information (as done for QRS detection algorithms) L69: Reference to MIT-BIH database should be entered L116-L117: Are the size of the sliding window (2000 samples) and the step (1600 samples) set according to the length of the records in MIT-BIH database or have they been set experimentally? Is it possible to change these parameters and compare the performance of the algorithms under different conditions? L134-L135: What do you mean? Is the noise detection method applicable only to MA or even to BW (as cited in L131)? The sentence is misleading and should be rephrased. L164 and L176: How was the threshold determined? Is it valid for all records types? More details should be added. L209: Why is β=1.5? The choice should be motivated. L252: “mean” instead of “men”? L259-261: The meaning of the “x” (indicated in Figure 14 – record 108) should be explained. L281-L282: What could be the cause of a lower DER than other methods evaluated on the same database? In contrast, SE and +P are similar as the authors have noted. Why? The hypothesis should be included as discussion of the results. L303: Is it possible to compare the average processing time with the methods in Table 2? L311: NSTDB is for? The full name should be introduced before the acronym. L319: “of for”.? L333: “form” instead of “from”?. L340: How many people have been analyzed? Information should be added. L341: Description of Figure 18 is wrong (equal to Figure 15). L372: A reference should be added to support the sentence. Discussion and Conclusions should be improved.

Comments for author File: Comments.pdf

Author Response

The paper is interesting as well as the application in real-time ECG monitoring.
I have some comments and questions for authors that can improve the overall quality of the paper, the readability and the understanding for the readers.

1. In my opinion, abstract contains too many details on methods that are not needed in this section. I think that authors should modify it and “balance” its structure (background, objective of the study, methods, results and conclusions) for greater impact on the reader.

Sol) As you pointed out, we modified the abstract.
Recently, with active development of wearable electrocardiogram (ECG) devices such as smart-bands or portable ECG devices, efficient ECG signal processing technology that can be applied in real-time has been actively studied. However, a wearable ECG device is exposed to various noise situations, thereby reducing the reliability of detected R point or QRS interval. In addition, as early warning techniques in healthcare systems have been studied, real-time ECG signal processing techniques have become very important in wearable ECG devices. In this paper, we propose an efficient real-time R and QRS detection method using two kinds of first-order derivative filters and a max filter to analyze ECG signals measured from wearable ECG devices in real-time. The proposed method detects R point and QRS interval in units of sliding window for real-time processing and combines the detected R points in each sliding window. Also, the reliability of the detected R points and RR intervals is examined through noise region analysis using histogram characteristic of sample point. The performance of the proposed method was verified by the MIT-BIH database (DB), CYBHi DB and real ECG data measured from the developed wearable ECG patch. The proposed method achieves Se=99.80%, +P=99.80%, and DER=0.36% against MIT-BIH DB. In addition, the proposed method enables accurate R point detection and heart rate variability (HRV) analysis even in noisy ECG signals.

2. L43-L52: some additional references on available technologies could be inserted to improve the state of the art information (as done for QRS detection algorithms)

Sol) We added the following 4 references for L70-76.
9. Dimiter, V.D. Medical internet of things and big data in healthcare. Healthc Inform Res. 2016, 22, 156–163.
10. Medtech and the internet of medical things. Available online: https://www2.deloitte.com/global/en/pages/life-sciences-and-healthcare/articles/medtech-internet-of-medical-things.html (accessed on 24 Sept. 2019).
11. Madias, J.E. A proposal for monitoring patients with heart failure via smart phone technology-based electrocardiograms. J. Electrocardiol. 2016, 49, 699–706.
12. Android app for patient monitoring. Available online: https://www.amrita.edu/center/awna/research/patient-monitoring-app (accessed on 24 Sept. 2019).

3. L69: Reference to MIT-BIH database should be entered

Sol) We added the following reference for L108.
2. MIT-BIH Arrhythmia Database Directory. Available online: https://physionet.org/physiobank/database/html/mitdbdir/intro.htm#annotations (accessed on 7 Aug. 2019).

4. L116-L117: Are the size of the sliding window (2000 samples) and the step (1600 samples) set according to the length of the records in MIT-BIH database or have they been set experimentally? Is it possible to change these parameters and compare the performance of the algorithms under different conditions?

Sol) The following description has been added for L126-133.
The setting for sliding window is set considering bradycardia. To get a successive R peak sequence even for bradycardia (less than 60 heartbeats per minute), the length of sliding window and overlapped interval between sliding windows should be at least 3 secs and 1 sec respectively. In the proposed method, the size of sliding window is set to 2000 samples (about 5.6 secs) based on MIT-BIH DB (360 samples/sec), and the step size of sliding window is set to 1600 samples. This means the overlapped interval between sliding windows is 400 samples (about 1.1 secs). So the length of sliding window and overlapped interval can be adjusted according to sampling rate of measured ECG signal.

5. L134-L135: What do you mean? Is the noise detection method applicable only to MA or even to BW (as cited in L131)? The sentence is misleading and should be rephrased.

Sol) The point you pointed out was corrected as the following for L146-154.
ECG signals measured in a measurement room or at stable posture are relatively clean, but ECG signals measured under motion may contain various noises.
The proposed method is basically robust to BW because it uses the product of two kinds of derivative filters. In addition, the processing of sliding window unit can minimize the influence of false detection by strong noise. The proposed noise detection method is designed to cope with BW and EMA.

6. L164 and L176: How was the threshold determined? Is it valid for all records types? More details should be added.

Sol) The following explanation is added in the paper for L184 and L197.
(L184) This threshold was experimentally obtained using the difference of the sum of respective vertical histograms for the MA signal of NSTDB [33] and the ECG signals of MIT-BIH DB [2].
(L197) This threshold is applied to the normalized max filter result, which is the final stage of QRS interval detection, and applies to all records of MIT-BIH DB.

7. L209: Why is β=1.5? The choice should be motivated.

Sol) The following explanation is added for L231.
The coefficient value is set to search for negative R or PVC below the average of ECG signal in the sliding window.

8. L252: “mean” instead of “men”?

Sol) We corrected the word for L274.

9. L259-261: The meaning of the “x” (indicated in Figure 14 – record 108) should be explained.

Sol) The following explanation is added for L282.
In the figure, x in Record 108 means non-conducted P-wave (blocked APB). It represents there will not be a QRS complex following.

10. L281-L282: What could be the cause of a lower DER than other methods evaluated on the same database? In contrast, SE and +P are similar as the authors have noted. Why? The hypothesis should be included as discussion of the results.

Sol) The following explanation is added in 4. Discussion (for L392-401).
In the performance comparison of section 3.1, the proposed method has the third lowest DER. The methods with the lowest DER and the second lowest DER are [28] and [19] respectively. However, since those methods do not have noise detection function and post-processing for R points and RR intervals contaminated by noises, it may be difficult to calculate reliable HRV. The reason why the proposed method has similar Se and + P is because the number of FP and FN is similar. This means that the existing methods may obtain biased FP and FN result as they are processed in the Record unit of the MIT-BIH DB, while the proposed method achieves FP and FN with uniform distribution due to the effect of increasing the amount of data by processing in the sliding window unit. Therefore, we can assume that the proposed method will have relatively uniform FP and FN regardless of the kind of experimental data.

11. L303: Is it possible to compare the average processing time with the methods in Table 2?

Sol) I’m sorry. We attempted to compare the average computation time with the existing methods in Table 2. The code in [14], [15], and [22] is published below on the github web-site, but the code for the rest methods is not published. It is also difficult to quantitatively compare due to the software mismatch of that codes on github (Python version) and our code (Matlab version).
https://github.com/luishowell/ecg-detectors
https://github.com/danielwedekind/qrsdetector

12. L311: NSTDB is for? The full name should be introduced before the acronym.

Sol) We added the full name and reference in section 2.2 and section 3.3.
(L138 in section 2.2) The ECG noises are extracted from MIT-BIH Noise Stress Test DB (NSTDB) [33].
(L345 in section 3.3) the NSTDB [33]

13. L319: “of for”.?

Sol) Thank you for that point. We removed the "of" for L353.

14. L333: “form” instead of “from”?.

Sol) We revised the passage as follows:
(L366) Analysis of Actual Data Measured by Wearable ECG Device

15. L340: How many people have been analyzed? Information should be added.

Sol) The sentence has been revised as follows for L371-374:
The 10-minute ECG data at fixed posture and walking for 36-year-old man without heart disease were measured. Figure 19(b) and 19(c) show the 1-minute length for 2 situations. Also, the R points on the measured ECG signals were detected by the proposed method.

16. L341: Description of Figure 18 is wrong (equal to Figure 15).

Sol) We revised as the following.
(L375) Figure 19. (a) developed wearable ECG patch; R point detection result at (b) fixed posture and (c) walking.

17. L372: A reference should be added to support the sentence.

Sol) We added the following references for L413-415.
That is, developing a technique that measures the reliability of detected R or calculated RR intervals is just as important as developing a QRS detection algorithm with higher detection rate [36,37].
36. Eguchi K.; Aoki R.; Yoshida K.; Yamada T. Reliability evaluation of R-R interval measurement status for time domain heart rate variability analysis with wearable ECG devices. Conf. Proc. IEEE Eng. Med. Biol. Soc. 2017, Seogwipo, South Korea, 11-15 July 2017; pp. 1307-1311.
37. Eguchi K.; Aoki R.; Shimauchi, S.; Yoshida K.; Yamada, T. R-R interval outlier processing for heart rate variability analysis using wearable ECG devices. Advanced Biomedical Engineering 2018, 7, 28–38.

18. Discussion and Conclusions should be improved.

Sol) We revised the discussion and conclusions.
L392-401 is added in discussion. The conclusion was revised.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors replied to my questions and comments, improving the readability and overall quality of the manuscript, including new references as required. The manuscript has been revised, clarifying some key points in the description of methods, results and conclusions. For this reason, The revised manuscript can be accepted in present form.