The Application Mode of Multi-Dimensional Time Series Data Based on a Multi-Stage Neural Network

Round 1

Reviewer 1 Report

I found this manuscript particularly hard to understand. The introduction section fails to clearly explain the motivations of this study. It is unclear what problems the proposed approach tries to solve, what innovations were offered in comparison with precedent works, and how well the proposed approach solves the problems.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors propose an approach based on multi-stage deep learning to deal with the feature learning of electronic medical data in reverse events risk prediction. In particular, the method incorporates both the medical use trajectory and the patient information to predict adverse events of concurrent medical use. Some comments are below.

- Table 1, and in general the introduction, should include a broader literature considering also more complex and ephemeral conditions to better understand and motivate the importance of the contribution, such as:

- https://doi.org/10.1016/j.envint.2019.02.013

- https://doi.org/10.1016/j.csl.2021.101298 

- https://doi.org/10.1016/j.compbiomed.2021.104949

- https://doi.org/10.1109/JPROC.2018.2791463

In particular, including not only methods based on deep learning, but also machine learning that performs well with EHR.

- The authors should clarify what "medical use trajectory" means. A definition is missing and the reading is not comfortable with that. For this reason, the contribution at the end of the introduction is not very clear.

- Line 102, is Figure 2 correct here?

- Line 130, where is the reference HINTON [19]?

- Line 139, “However, the structure of AE is so simple that it only has three layers of neural network” but encoding and decoding may have many more layers. It is not clear if this is a design decision or a general statement. In the former case, the authors should clarify why they did not try a complex architecture, in the latter a reference is needed.

- Line 151. “Figure 1 illustrates” the contribution is not very clear.

- In the results section, the images are small and grainy and the labels are unreadable.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors need to do the following corrections / modification to this articles.

·         In abstract – acronym should not be used and need to be replace

·         Fig 1: Deep LSTM-AE – the figure text not visible

·         Equations and equation numbering to be checked and placed properly.

·         The authors need to do English grammar checking and editing.

·         The results are supporting to their research findings.

·         They need to do minor corrections and submit it.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

In the study entitled “The Feature Learning Method of Electronic Medical Data Based 2 on Multi-stage Deep Learning Models”, Wang et al. have proposed a multi-stage deep learning-based feature learning method (MDFL) for feature learning of high-dimensional electronic medical data in adverse events prediction. The topic is of good importance, as feature learning is one of the most important phases for accurate prediction. However, there are a few issues in the manuscript that needs to be addressed. Also, various important details are missing in the manuscript, which can help the readers to understand the method and to reproduce the results.

1. The language of the manuscript needs to be reviewed by a native language expert, as there are various grammatical and sentence structuring issues.

2. Abstract lacks to provide insights into the results of the study.

3. Introduction lacks a discussion on state-of-the-art methods available for Bioinformatics, as this area has advanced since the last decade and such signs of progress should be discussed here.

4. Details of the model architecture and their hyperparameters are missing. Such details are crucial for the reproducibility of the method.

5. Which feature set performed well as compared to the others? Authors also need to compare their method with existing feature engineering methods for medical records.

6. Results should be updated accordingly, after a thorough comparison. 

I would recommend the author address all these issues, as, without addressing them, the manuscript lacks substantial information.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

The authors have addressed all the comments. I recommend the manuscript for publication.