BioECG: Improving ECG Biometrics with Deep Learning and Enhanced Datasets

Round 1

Reviewer 1 Report

Dear authors,

The manuscript “BioECG: improving ECG biometrics with Deep Learning and enhanced datasets” written by Paloma Triado-Martin et al reports Deep Learning-based ECG biometric recognition. However, there were major and minor issues that were not properly addressed by the authors in this manuscript.

 

Major Comments:

1. A clear explanation for D and V from the selected database is needed. In addition, as a case of ‘D’, it is hard to understand the correlation between ‘one-day and two-day enrolments’ and ‘represents the day’ (it is the crucial point to mislead the Table). The authors should clarify this point.

 

2. In Table. 3, the authors selected set samples (Hd) using set proportion (d) and divided training set (Htrain) and validation set (Hv) with the ratio of 4:1. However, in the case of 0.5 (set proportion), the ratio of the training set and validation set is different, which raises the question about the accuracy of results in Table. 4(a) and Table. 5(a). The authors should address this point.

 

3. When the author extracted many distinctive features using CNN, the specific explanation of the composition of each layer and output result are essentially required. The author mentioned the size of the kernel, filter, and strides, which is able to predict the size of output, however, missed information of CNN (i.e., number of layers and used filter) degraded the readability of the manuscript. The reviewer recommended to the author to newly include the explanation of CNN in detail.

 

4. In the pre-processing stage, the effect of the butterworth filter for the result is required. Although the reviewer also recognized the pre-processing is essentially needed for signal data, the performance comparison between other grades or filters and the fifth-grade butterworth filter is missed. This issue is needed to be addressed to validate the reason for utilizing the fifth-grade butterworth filter.

 

5. The author used EER as a quantitative factor to evaluate the performance of classification and BioEEG. However, the description and equation for deriving the EER. The explanation of the deriving process or equation of EER is required to enhance the readability of the manuscript.

 

6. For the comparison of Table. 9, on the contrary to 1 enrolment is used at previous study, the author utilized 2 enrolment for the proposed method. Therefore, it is not sufficient to mention ‘could provide more insight into this modality’ (line 478).

 

Minor Comments:

1. In the introduction part (Lines 77 ~ 84), the author mentioned the disadvantages and limitations of other classification models (i.e., SVM, KNN) for this study. Please include the references to support each sentence to enhance the reliability of the author’s statements.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

The manuscript presents an application of BioECG neural network for human verification and identification with tuning of enrolment length and considering more complex and realistic biometric environments, such as changes in body position and heartrate. Authors report promising results when body position is not changed, especially when the enrollment is done in two different days.

I have the following specific remarks, questions and recommendations:

• Section ‘Introduction’ should be extended to present more attempts of human verification and/or identification by means of ECG signals. There are a lot of papers published in the last 10 years that deal with application of the ECG for biometrics in terms of: database suitability (Merone et al, “ECG databases for biometric systems: A systematic review”; Jekova I et al, “Human identification by cross-correlation and pattern matching of personalized heartbeat: Influence of ECG leads and reference database size”); variation of the ECG signal within one person and between humans (Jekova et al, “Intersubject variability and intrasubject reproducibility of 12-lead ECG metrics: Implications for human verification”); presentation of different techniques for verification or identification (Tan et al, “Toward Improving Electrocardiogram (ECG) Biometric Verification using Mobile Sensors: A Two-Stage Classifier Approach”; Islam and Alajlan N, “Biometric template extraction from a heartbeat signal captured from fingers”; Krasteva et al, “Biometric verification by cross-correlation analysis of 12-lead ECG patterns: Ranking of the most reliable peripheral and chest leads”; Arteaga-Falconi et al, “ECG Authentication for Mobile Devices”, etc.).
• Section ‘Materials and methods’:
• Details on the QRS segmentation described in subsection ‘3.1. Pre-processing’ are necessary. A figure would be welcome. At this point of the manuscript it is not clear that +/- 100 ms around the R-peak are considered, which provokes questions.
• The description of the input data in subsection 3.2.1 is not clear. A possible reason could be related to the fact that the ECG database used in the study is still not described. In my opinion the database description should precede the methodology description. The authors should explain that for each user (U) equal number of signals (Ns) are extracted and from each signal equal number of segments are extracted. Otherwise, the equation Hdb = Hs * Ns * U is not correct.
• Subsection ‘3.3. Database’ – There is a missed reference (“…ISO 19795 requirements [? ].”). How did the authors succeed to record ECGs with exactly 130 bpm for each participant in the experiment? Related to this, I would recommend the use of “hear rate of 130 bmp” instead of “heartbeat frequency to 130 …” in the entire manuscript.
• Subsection ‘3.4.1. Input data’ (subsection of ‘3.4. Modelling) - The sentence “Each user provides N = 50 per signal, and 5 signals per visit so Ns = 5, concluding in 250 samples per user. It finally translates into an available data dimension of (250, 200) per user.” is not clear. How does this information correspond to Table 1? Is “user” equivalent to “subject”. One and the same terminology should be used in order to avoid ambiguity.
• Subsection ‘3.4.2. Classifier’ (subsection of ‘3.4. Modelling) – on what subjects/users are the “several heuristic observations” done? Are they obligatory included in the training part of the data? Explain! Further it is written “The network is provided with values from 1 to 3” – values for what?
• Subsection ‘3.4.3. Output’ (subsection of ‘3.4. Modelling) – is (Htrain, U1) a matrix containing 0 and 1, or it contains values between 0 and 1? Explain the content of this matrix (rows and columns).
• Section ‘Experimental analysis’: I would propose the authors to extend the table’s captions, so that they contain sufficient information for understanding the presented results. This would facilitate the reading and understanding.
• The authors should read very carefully the manuscript and correct some language errors.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 2 Report

The authors have considered the most important remarks and suggestions in my first report and the manuscript could be published in its present form. However, I would drive the authors' attention to the fact that they should address each reviewer's remark separately and highlight the related changes in the text. Now they have selected to what they want to answer and correct, which is not a good approach.