Attention-Oriented CNN Method for Type 2 Diabetes Prediction
and Lijuan Shi 2,3,5
Shreya Chakrabarti
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsSome questions:
1) Is there no more recent related work, e.g. 2023?
2) Aren't there too few examples in the PID dataset for training a neural network?
3) What are the characteristics selected by the ANOVA and LP method?
4) Why use a CNN to select the most relevant features if the ANOVA and LP methods have already done this task?
5) What is the data input format for the CNN network? It seemed to me to be a vector, so why use a CNN, which receives feature matrices (like an image)?
6) Why use an ANN if you have already used a CNN?
7) Why didn't you use more current and better performing Machine Learning algorithms such as XGBoost and LightGBM?
8) What is the main objective of the work, to make a prediction or an inference?
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsOverall, the paper evaluates a few different ML algorithms on Diabetes prediction and presents reasonable conclusions in terms of the most performant algorithm. I have the following questions/suggestions for authors:
1. It will be good to add explanations on why certain modeling choices were made, e.g. why did the authors use mean to fill missing values, why did we use ANOVA, etc.? We should particularly explain these choices in the context of the structure of the data being used.
2. It is unclear why the authors used ANOVA and LR. The number of variables is already quite low and there is no need to do feature selection. In case this is for model explanability, it will be good for the authors to explain the findings in clinical terms and add a validation by clinician.
3. SE already exists in prior work and CNN is a commonly used ML model. What is novel about the SECNN in this paper? Specifically, what is the novelty being introduced by the authors?
4. How will we ensure that these results generalize to other diabetes datasets?
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for Authors
Dear authors,
Thank you very much for your contribution. Please find below some comments that may help you to increase the soundness and the readability of your manuscript.
1. Line 27: “Diabetes (…) faces huge challenges”. While I understand what the authors want to say, the phrasing is not correct. It is the fight against Diabetes that faces huge challenges, rather than the disease itself.
2. Several claims are made from lines 29 to 41 without any reference being made. I suggest the authors to add references that corroborate what is being said.
3. Line 33: “which we call “three more and one less””. The authors need to clarify to whom does “we” refers to.
4. Line 35: “T1DM (…) represents one of these conditions”. The authors already said that, consider maintaining only new and relevant information on this phrase.
5. Line 45: “It is often (…), arteriosclerosis”. Please consider rephrasing.
6. Lines 51-58: Please carefully confirm every statistic presented in this paragraph.
7. Lines 59-61: The sentences are repeated.
8. General “Introduction” comment: Add a reference for each claim made.
9. Lines 101-102: I recommend that the authors introduce the acronyms before using them.
10. General “Related Work” comment: This section is composed of a big paragraph that it is difficult to follow. The authors present an enormous amount of results without stating to what they refer. For example, the reader cannot understand how is the classification being done or what type of features are being selected. Since a screening of the literature is important to evaluate the soundness of the manuscript I recommend that the authors rearrange this section so that it is clear what has already been done in this field and what gap in the literature is being filled by their research.
11. Line 139: “the data set used in this article, the framework proposed in this article”. Please consider rephrasing so that you do not repeat “in this article”.
12. NHANES dataset description is very confusing. The authors state thar “this dataset contains 51 different variables”, however, they also state that “in this article, we used a set of 25 different variables”. How did the authors chose which variables to use and which variables to reject?
13. Still regarding the NHANES dataset, the authors state that the 25 used variables are “essentially independent” and then provide a table with descriptions of categorical variables that are “dependent”. There is a clear confusion regarding the term “independent”. The authors need to clarify what they mean by “essentially independent” and also need to state why are the variables on Table 2 “dependent”. Moreover, Table 2 only shows 7 variables, what about the other 18?
14. It is imperative that, in section 3.1, the authors not only state exactly which variables were used in model development but also include a table with the number of instances of each class (diabetes/no diabetes).
15. In section 3.2.1, the authors state that the use of mean(x) to fill missing values “can maintain the continuity of the data distribution without importing outliers”. I suggest that the authors clarify what is meant by “importing outliers” because the mean is sensible to the presence of outliers.
16. Line 183: “The reason for outliers include errors, input errors, or extreme events”. Please consider rephrasing.
17. Line 189: “We set the threshold to 1.5” – What was the rationale behind this choice?
18. Feature importance analysis subsection: The authors state to use ANOVA analysis, did the authors tested the data for normality? Moreover, when using LR, the authors use a p-value of 0.05 but multiple comparisons are being made. The authors should clarify if any correction was done to account for these comparisons.
19. Assigning Class Weights subsection: As the authors state, there is a serious problem regarding class imbalance in this work. It is important to understand how the class weights were assigned. The authors only point the use of a Python library, that is insufficient.
20. Lines 240-258: The discussion regarding CNN is outside the scope of this paper. I recommend that the authors remove this paragraph and just point the reader to a reference paper. I think this modification will improve the readability of the manuscript.
21. Lines 259-279: All the information regarding CNN architecture needs to be summarized on a table.
22. As said for CNN, the information given from lines 300 to 387 can be removed.
23. The authors explain and give formulas for the calculations of each performance metric except AUC, why?
24. Lines 439-440: “We identified outliers (…) and used median replacement”. So, did the authors use the mean or the median, for replacement? Please clarify.
25. Figure 8: There is a row missing in (d). Also, even with zoom, it is very difficult to evaluate figure (c) and (d).
26. Lines 460-461: “the correlation coefficients (…) are significantly improved”. Which statistical tests were performed to assess significance.
27. Tables 4, 5: The quality of these tables needs to be vastly improved specially regarding the number of significant figures used in F-score and P value
28. Table 6: Same comment as for the other tables, numeric values cannot be presented this way.
29. Line 497: “performed better in performance.”
30. Tables 7,8: Please improve the division of the table concerning “K value” it is difficult to read. Also, there is a number in bold in table 8.
31. Figure 10: Graph title should be changed to a more representative one.
32. Lines 503-513: Please carefully check every value presented here.
33. Line 531: The authors should carefully clarify if they are analyzing the increase in percentage or in percentage points.
34. The authors should explore deeply the works in the literature and compare them with their results, without that, is impossible to evaluate the impact and soundness of the manuscript.
Comments on the Quality of English Language
Moderate editing of English is needed, as stated in detailed comments.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
Comments and Suggestions for Authors
Dear authors,
Thank you very much for your answers to my comments and suggestions, however, some clarifications are still needed.
1. The authors stated in their response to my first comments: “we did perform a normality test” – to assure that ANOVA can be used in the feature importance step. Nonetheless, this information (and its results) was not included in the manuscript. Why? Since the work is highly based in the selected features, the paper cannot be accepted until the authors clarify, and prove, that normality was assessed.
2. “we found that the Pearson correlation coefficient between variables in the data set was significantly enhanced”. Do the authors have a p-value to corroborate this statement?
Comments on the Quality of English LanguageNo comments.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
