A Deep Recurrent Neural Network-Based Explainable Prediction Model for Progression from Atrophic Gastritis to Gastric Cancer

Round 1

Reviewer 1 Report

Dear Authors,

Undeniably, gastric cancer is an important research topic and data science is a much needed aid in attempt to ameliorate its negative impact on population health. In that general regard, your study is an interesting one, however I detected some serious flaws which prevented me from accepting it in present form. 

First of all, you should consider extensive editing of English language. Second, you methodology need more explanations. For example, you clearly state that the data is seriously imbalanced, but then you proceed treating it as balanced because I quote "of maintaining the incidence 189 of gastric cancer in atrophic gastritis patients". This is either wrong or not sufficiently well explained, because balancing of the dataset while training is much needed in your case, while test set should remain unbalanced. This flaw can be clearly seen from your results since you predict correctly around 50% of gastric cancer patients in the minority class, while the rest are false negatives, which is not a particularly good result. 

Also the three groups in clustering results and the chi-square used to evaluate feature importance are not particularly clear. Please look at interpretable machine learning models such as interpretml (https://github.com/interpretml/interpret) in order to test your results and provide more useful benchamrking.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Τhe authors predicted patients with atrophic gastritis who were at high risk of becoming gastric cancer patients and analyzed some of their characteristics. For this purpose, they used DeepPrevention, which is a deep recurrent neural network (deep RNN)-based prediction model to find risk factors and detect high risk for progression to gastric cancer from atrophic gastritis. Also, they used K-means clustering to predict the patients at high risk for gastric cancer. The proposed model achieved an area under the receiver operating characteristics curve (AUC) of 0.84.

The article topic is tοο crucial for human life and, the authors come up with results that cοuld be used in future research.
The article has a proper structure, and there is a flow between the sections. 
The paper is well-organized and described the details adequately. 
The authors should remove all issues in terms of English writing. Ιt requires moderate proofreading.
The bibliography is sufficient and up-to-date.
The article has good prospects, but some changes and reinforcements need to be made.

1)I suggest the authors reduce the length of the abstract. Such a detailed description is not necessary.
2)I would like the authors to make a more detailed technical description of the proposed DeepPrevention model. Describe the model algorithmically step by step in the form of pseudocode.
3)Give more details about the implementation environment of your experiments(source code, programming language, tools, frameworks, links). It is only noted that they were conducted in the Windows 10 environment.
4)Τhe authors need to present more graphic illustrations to strengthen their model (namely, AUC ROC). They should be accompanied by appropriate commentary.
5)In Figure 4. (a) what are the X, Y axes? Also, the resolution of figures 3 and 4. (b) needs correction.
6)The doi of references [7],[8],[9],[17],[22] and [23] are not accessible.
Moreover, I could not find and read the reference [19].
7)The Conclusions section needs a better connection with the simulation results. The authors should give numerical details about the deep model performance.
8) Please, in the discussion section, compare your model performance with other related works on this topic. You can summarize them in a table, for easy comparison. 
9)As a future direction, the authors plan to develop variations of the proposed deep recurrent neural network model to achieve higher sensitivity. Under what conditions cοuld these be achieved?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

The topic is very interesting.  My comments are as following.

1. Need to give more details on the proposed “deep recurrent neural network-based prediction model”,
2. Need to provide more information about the Chi-square test in Table 4. Which part of the data are used?  Why gender, body mess are not included?  Why family cancer is included?  How did you choose the variables?
3. Try not to repeat. For example, formulas (1) and (2) repeat the information in the last paragraph of page 5.
4. What the purpose of Figure 2? What are the component 1 and 2?
5. Figure 1 can be replaced with a couple of sentences.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Dear Authors,

You have improved the quality of your manuscript significantly and due to the importance of the topic you research, which is acting on prevention rather then predicting outcome, I can conclude that your work is of sufficient quality to be published (however this will off course depend on the final judgement of Editor-in-charge). My final suggestions to you are following: When you used InterpretML and observed that the important features affecting the prediction of gastritis patients differed from patient to patient, this deserves more detailed study and explanation! 

Regarding the future course of your study, my suggestion is to certainly include information on h.pylori infection and to consider gut microbiome 16S rRNA sequencing, which I strongly believe would provide big boost to your model's predictive power. I am still not totally convinced on your strategy for handling imbalance, but you did explain your approach and you worked with what you have available and I can live with that.

Good luck!

Author Response

Please see the attachment

Author Response File: Author Response.docx

Reviewer 2 Report

I have no additional remarks on the revised version.

The authors have addressed my concerns.

Thank you very much.

Author Response

Thank you for your valuable comments and suggestion.

Reviewer 3 Report

Thanks for the detailed responses.

Author Response

Thank you for your valuable comments and suggestion.