Investigation of Methods to Extract Fetal Electrocardiogram from the Mother’s Abdominal Signal in Practical Scenarios
, Tai Le 1,†, Amir Mohammad Naderi 1,†, Quoc-Dinh Nguyen 2,†, Brandon Tiang-Yu Kuo 1, Tadesse Ghirmai 3, Huy-Dung Han 2, Michael P. H. Lau 4 and Hung Cao 1,4,5,*
Round 1
Reviewer 1 Report
The authors describe several analysis algorithms for recovering the fetal heart rate from an abdominal ECG from the mother. This is important work that is for the most part well-written, and is helpful to the community as it directly compares several filtering approaches. I recommend the following revisions, specifically emphasizing the need to add relevant citations in the introduction and methods.
- Must include more references in the introduction, with the second paragraph especially lacking relevant sources to back claims and context.
- I did not see QRS defined. Please add an explanation.
- The drawbacks of CNNs are not adequately addressed. The memory issue is somewhat irrelevant, for simple networks, like ones that may analyze 1 dimensional data, can be executed easily (in near real-time) once trained. A similar argument can be made “for the need for many hidden layers” – this is simply not a big deal, computationally speaking, for the interpretation of low-dimensional data like describe in this manuscript. The main drawbacks would be the need for training data (As stated) but also the tendency to over-fit to training sets and noise environments. I would recommend updating this discussion to more accurately discuss this approach, especially because it has had much success in the field of biomedical sensing and imaging.
- ‘aECG’ acronym should be defined in line 109-110.
- Equation 2 should define ‘w’ that is under the summation or set it equal to 5.
- Section 2.3 should have improved organization. Sub sections should be mentioned in the order they are written. Additional text under the 2.3 heading could help orient the reader to the following subsections, as there are many.
- All sub sections should cite or mention the specific functions or code bases used to execute the algorithms.
- Section 2.3.2 – last sentence “…which is illustrated below.” must reference a figure/table/ or algorithm.
- Must have a citation for subsection 2.3.4 unless the algorithm is devised by the authors. If it is devised by the authors it should be stated in this subsection.
- There are two copies of the same sentence at the beginning of section 2.3.5. Please correct.
- Section 2.4 ‘RR series’ must be defined.
- Figure 3 needs figure caption describing (a)-(d).
- How is figure 3d made exactly? Which ICA method is used to filter/alter the signal? Please explain.
- What is the “2.” In Equation 6? This might confuse readers unless it is defined. Please correct.
- The paragraph starting with line 329 should better contextualize what this timing means. How much data (in terms of the time of the zECG) is processed during this time?
- The information included in the first paragraph of the conclusion should be moved to the introduction. This information should also include citations and specify if there is a direct fECG measurement or if these systems similarly use indirect and data filtration methods for obtaining the fECG signal. The authors must be specific when comparing hardware and software solutions, as their implemented algorithms could be employed on the existing hardware solutions and therefore exist together instead of stand-alone as a solution.
- Additionally, the last paragraph should better explain that the filter methods used to obtain the fECG were processing data from mECGs.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The presented study deals with the comparison of different fetal ECG (fECG) extraction algorithms and consequent QRS detection in fECG. This is an actual problem, which has not been reliably solved yet and attempts in this field are worth to be published. However, I have the following critical remarks and recommendations that should be strictly addressed before positive decision for publication of the manuscript:
1) The abstract should be extended to present the achieved results – numerical values of F1 at least for the best performing method. The authors could decide to provide also min-max range if it is possible to address some particular methods in the abstract.
2) Section ‘Introduction’:
- I wonder if mobile health (m-Health) shouldn’t be abbreviated at the beginning of the Introduction;
- Abbreviate CDC;
- The Introduction could be extended by addressing more articles from the CinC Challenge 2013 and especially those published latter in the special issue of Prysiol. Meas. 2014.
3) Section ‘Materials and methods’ – What do the 4 noninvasive abdominal signals contain? Four aECG signals acquired from different points or something else (e.g. is there a channel containing only the maternal ECG). Provide the necessary explanations. Concerning the database – have you used separate training and test datasets? If not, this should be noted as a limitation of the study. Discussion about accuracy drop on unseen aECG signals should be included.
4) Section ‘Materials and methods’, Subsection 2.1 ‘Extended Kalman Filter’:
- Abbreviate EKF at the name of the subsection. Now, it is not absolutely clear where this abbreviation EKF comes from;
- The last paragraph in subsection 2.1 in fact describes the application of EKF for filtering the mECG (“Based on this dynamical model, a few algorithms that employ EKF for the extraction of fECG from abdomen signals have been proposed. One such algorithm used sequential EKF algorithms that first filter out the mECG signal, assuming the mixture of the fECG and noise as a Gaussian noise, followed by removing the mECG signal by subtracting it out from the abdominal signal and another EKF that filters out the fECG from the noise.”). This is one of the estimated methods in the manuscript, therefore, it should be described in more details (schematically or as a clear sequence of actions with clear input and outputs).
5) Subsection ‘2.2. Template subtraction (TS)’:
- What is TSc?
- Correct “last-mean square (LMS)” to “least-mean square (LMS)”;
- I have the same critical remark as for the previous subsection. .”). This is one of the estimated methods in the manuscript, therefore, it should be described in more details (schematically or as a clear sequence of actions with clear input and outputs). Moreover, all variables should be written in italic shrift, so that they are recognizable in the text;
- Equation 4 – It is not obvious that the ECG cycles ‘m’ and ‘t*a’ are subtracted sample by sample. I recommend ECG to be included in the equation. Details about the method are necessary (e.g. is cycle alignment necessary, how is t done, etc.)
6) Subsection ‘2.3. Independent Component Analysis (ICA)’:
- What is TSc?
- Check this statement: “the sample value of xj is denoted by xj”. It is not clar.
- In subsection ‘2.3.5. TS and ICA combination’ – The following sentence is duplicated “The combination of different methods could yield higher performance.” Moreover, I would recommend (1) TS-ICA; (2) ICA-TS; and (3) ICA-TS-ICA to be followed by their a little bit extended explanation. E.g. (1) TS-ICA, which is …; (2) ICA-TS, which is …; and (3) ICA-TS-ICA, , which is …
7) Subsection ‘2.4. Peak fECG detection’:
- In the practical case you will not have any annotations of the places of the fetal QRS complexes and you will not be able to apply the following: “This detection was performed independently on all four output channels after doing extraction and the best RR series which was close to labeled peak fECG annotations was selected.”
- Figure 3 – I suppose that it is presented for the best channel. As I mentioned above, in practice you will not be able to select this best channel, since you will not have on your disposal any fetal QRS annotations. Provide also the signals in the remaining 3 channels, so that the reader can judge what will be the performance in practice. Moreover, use one and the same y-scale (at least in subplots a, b, c). Thus, the correspondence between a and b, c would be visible.
- If the accuracy results (section R) are presented for the best channel – provide results also for the remaining 3 channels or justify the use of only the selected channel. Normally, the selection of channel should be done on a training dataset and after that in the test dataset only this channel should be considered. However, in this study the data is not separated for training and test. What is sure is that one channel should be selected and the fECG extraction and subsequent QRS detection should be done on the selected channel foe all ECG recordigs.
8) Subsection ‘2.6. Comparison schemes’:
- Figure 5: again the same requirement as for Figure 3 - use one and the same y-scale for the different subplots.
- “A statistical analysis which assesses the detection of the QRS complex waveform of the fECG was performed by comparing the waveforms of the dataset and the extracted fECG.” – comparing the waveforms leaves the impression that you search for morphological differences in the waveforms (via sample by sample subtraction or correlation assessment, or some other approach). In my opinion you just compare the positions of the detected and annotated QRS. Is this right? Explain what have you done exactly and if you do not assess the differences in the QRS waveforms, correct the sentence in an adequate way.
9) Section ‘3. Results’:
- Figure 5: again the same requirement as for Figure 3 - use one and the same y-scale for the different subplots.
- The authors have not used separate training and test datasets and this inevitably leads to higher performance. All adjustments (lead selection, threshold values in the applied methods, etc.) should be done on a training dataset and after that tested on an independent test dataset. If not, discussion about accuracy drop on unseen aECG signals should be included in section ‘Discussion’.
10) Section ‘5. Conclusions’:
- The first paragraph in this section is not a conclusion from this study. It is more appropriate for ‘Introduction’.
11) The reference list is not in the correct format.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The authors have considered the remarks and recommendations and have provided answers to the questions in my report. The present version of the manuscript is suitable for publication. I have just one minor remark - I could not understand the following sentence, which I recommend to be rewritten:
“The fECG channel was chosen based on a smoothing indicator (SMI), which is the number of times the absolute value of the change in instantaneous heartrate more than 29 beats per minute and beat comparison measure (BCM) [16]. Finally, fQRS was detected using the Pan-Tompkin algorithm.”
