A Novel Algorithm for Automated Human Single-Lead ECG Pre-Annotation and Beat-to-Beat Separation for Heartbeat Classification Using Autoencoders

Round 1

Reviewer 1 Report

Comments to Authors (General)

·       The authors should cite more relevant recent works such as “https://www.sciencepublishinggroup.com/journal/paperinfo?journalid=170&doi=10.11648/j.ijpbs.20170206.12” and “https://medcraveonline.com/MOJABB/analysis-on-conversion-process-from-paper-record-ecg-to-computer-based-ecg.html”, and so on.

·       The authors shall have to amend some typing mistake in the manuscript. The following are some examples and it was found with red color in this comments.

In the line No. 417, “… to find the which ….” shall have to be amended.

In the line No. 502, “…heartbeats sue to the high …” shall have to be corrected.

In the line No. 517, “…results she reconstructed …” shall have to be modified.

·       The authors shall have to check the whole manuscript without typing errors.

·       The authors shall have to correct some sentences based on the technical English for international journal publication standards.

Comments to Authors (Specific)

·       The authors shall have to modify the section of “2. Materials and Methods” based on the additional key mathematical expressions for specific structure of the proposed system and give the discussions in detail.

·       The authors shall have to give the important algorithms for implementing the proposed system like “Algorithm 1” in page 10. All algorithms are necessary for full information in the manuscript.

·       The authors shall have to give detail discussions on Figure.6. The significant points shall have to be mentioned.

·       The authors shall have to recheck Figure.7 with exact numerical expressions in interpretation on the results.

·       Even though this proposed system was used novel algorithm on Autoencoders, the performance comparison with similar recent works in numerical or other values for the statistic table before discussions and conclusion section. The table 1 is acceptable according to the different data sources but authors shall have to give the performance comparison with other contributions based on processing time, delay time, system robustness, exactness.  

·       The authors shall have to mention the recommendations on analysis results for further extension in future.

·       The conclusion should be modified with expressions on proof of research problems in the manuscript.

Author Response

- General:
* The proposed papers are very much related to our paper and are now added to the cited references;
* The typos in the manuscript are checked and fixed;
- Specific:
* Section 2 regarding materials and methods is extended with the missing parts from the proposed workflow, and the exact details could be checked at the core of the manuscript.
* The important algorithm that needs all the details on how to be programmed is the pre-annotation part, however, the rest is explained as mathematical formats or as workflow depending on the case (for instance, the connection between the sub-functionalities is explained as a workflow, however, the detection of the threshold require the exact formula for calculating the discrete derivative and determining the used threshold). Every functionality is made in a standalone sub-module and its inputs/outputs could be related using the workflow;
* The discussion part was missing for figure 6, but now it is well explained and the limits of the detection are explained;
* The difference in the y-axis in Figure 7 resulted from the use of different datasets which were generated using different sensors/devices so that the voltage is different from one dataset to another. in fact, this is one of the main reasons to train a model to be able to exclude these differences and to be able to detect the correct shapes of normal heartbeats.
* The performance comparison with pre-existing solutions is added to the discussion (the same comment was also received from another reviewer);
* The conclusion is updated with the statements on proofs of results as well as further work.

Reviewer 2 Report

This paper presents a new algorithm for automated human single-lead ECG preannotation and beat-to-beat for heartbeat classification using autoencoders.

This paper provides an in-depth survey of similar work where neural networks have been used. However, when presenting the results, they should be compared in more depth with this state of the art, comparing qualitatively and quantitatively the results obtained. 

To validate the algorithm, different databases with noisy and non-noisy signals are used. And although this is the classical procedure, it would be very interesting to quantify in some way not only the level of noise, but also its characteristic (specifications) in order to be able to draw conclusions about the use of the algorithm on other signals.  That is why it would be very interesting to study how the algorithm behaves separately and jointly against: a) motion artefacts, b) when the signal is affected by important values of external interferences (60 Hz,...), c) electromyography signal and d) low frequency noise and drifts.  All this would make it possible to validate whether, beyond its application in databases, it is potentially useful for the processing of ambulatory measurements, which are increasingly present in today's telemedicine.

Author Response

* The performance comparison with pre-existing solutions is added to the discussion (same comment was also received from another reviewer);
* The validation of the results was already validated against low-level noise, as well as some intentional motion on the ambulatory sensor which was now explained more in details about the Figure 6. this was added to the current version of the manuscript.

Reviewer 3 Report

In this work, the authors have presented a Neural Networks based workflow for patient monitoring using the ECG data. The trained models show promising results for ECG pre-annotation and beat-to-beat separation for heartbeat classification. Authors need to incorporate the underneath mentioned points for a better and possible publication by the journal. I therefore, recommend for major revision.

1.       The abstract should be had the main achievement and emphasise the novelty, so please rewrite it considering these points. The research question has not been put forward clearly. Readers are not sure what specific problems you want to solve.

2.       The major objectives of the proposed work should be included in the introduction part.

3.       What is the motivation behind conducting the research. It is recommended to highlight major contributions of the work. Novelty of the work is not highlighted.

4.       The critical literature review should be presented to indicate the drawbacks of existed approaches, then, well define the main stream of research direction, how did those previous studies performed, which problem still requires to be solved and why is the proposed approach suitable to be used to solve the critical problem?

5.       The results appear to be too preliminary and incomplete for publication at the present time. Key results are poorly explained.

6.       Make comparisons with previous work in the field and include articles that have a clear statement of the improvements made to justify publication.

7.       There are several typos, e.g. Algorithm is not Figure. Axis and scale are missing in all the Figures.

Author Response

* The last part of the abstract is re-written in order to better explain the problem to solve and to strengthen the importance of our paper;
* The novelty is extended in the Introduction as adviced, especially for the pre-annotation process which was not well explained (this is from your questions 2 and 3);
* the main critical problem as defined is to provide a workflow to detect normal vs. abnormal heartbeats. this is why we started by stating the problematic in the related work and the introduction, then we presented the proposed solution with its implementation;
* the results part and the discussion is now extended with the advice of another reviewer, we added also the performance comparison with other pre-existing solutions together with our results which showed very good results compared to others (this is regarding questions 5 and 6);
* The present template from the journal does not provide a possible way for Algorithm insertion, this is why we decided to include it as a figure instead;
* The figures are updated to contain the axis legend.

Round 2

Reviewer 2 Report

The authors have correctly introduced into the article the changes and explanations that had been mentioned in the review process.