Next Article in Journal
Determining Optimal Sampling Numbers to Investigate the Soil Organic Matter in a Typical County of the Yellow River Delta, China
Next Article in Special Issue
A Graph-Based k-Nearest Neighbor (KNN) Approach for Predicting Phases in High-Entropy Alloys
Previous Article in Journal
A Study on Customized Prediction of Daily Illness Risk Using Medical and Meteorological Data
Previous Article in Special Issue
Object Detection Related to Irregular Behaviors of Substation Personnel Based on Improved YOLOv4
 
 
Article
Peer-Review Record

Use of Artificial Neural Networks to Predict the Progression of Glaucoma in Patients with Sleep Apnea

Appl. Sci. 2022, 12(12), 6061; https://doi.org/10.3390/app12126061
by Nicoleta Anton 1, Catalin Lisa 2,*, Bogdan Doroftei 1,*, Silvia Curteanu 2, Camelia Margareta Bogdanici 1, Dorin Chiselita 1, Daniel Constantin Branisteanu 1, Ionela Nechita-Dumitriu 1, Ovidiu-Dumitru Ilie 3 and Roxana Elena Ciuntu 1
Reviewer 1:
Reviewer 2: Anonymous
Appl. Sci. 2022, 12(12), 6061; https://doi.org/10.3390/app12126061
Submission received: 22 April 2022 / Revised: 10 June 2022 / Accepted: 13 June 2022 / Published: 15 June 2022
(This article belongs to the Special Issue Applied Artificial Intelligence (AI))

Round 1

Reviewer 1 Report

Line 94-95  the choice between larger layer with more neurons vs more layers largely depends on the problem itself. If you are stating a fact pertinent to your problem, you need reference to back up such claim.

Line 101-102, MLP and feedforward neural networks are not interchangeable terms.

Line 108-109, mse, r2, etc are for regression type neural networks.

Line 114, please consider reword the sentence.

Line 144, about 10% of data was kept out for test. This might not be sufficient to evaluate model performance. Also, with increase of number of neurons from 5 to 35, the number of estimation parameters grows quickly, and so is the need for more data to properly train the model.

The number of patient data is rather limited, i.e., 73 in total. Training a proper neural network model is usually data heavy. Have you considered other machine learning approaches?

Line 170, training a neural network model for 50,000 epochs without early stopping would likely cause overfitting to training data. Other than training epochs, there are other training parameters, learning rate, optimizer, loss function, etc. that need to pay attention to. Please discuss as appropriate.

Line 175, describe how input data were processed because the input variables do not share the same units/ numerical ranges.

Line 241 table 2. Are these MSE , R2, Ep values obtained from training data? As is apparent in the data, a more complex neural network model is preferred over simple ones. The trend is consistent and it is an indication that overfitting is present. It is also unlikely that both training and validation have the same outcomes, unless the optimal solution has not been found. Training performance always improve while validation performance will deteriorate at certain point.

Line 262, how did the author draw the conclusion that over thousands of data is needed based on fig. 10 and 11? Please explain.

Line 271-318, this section can better fit in the introduction or literature review section since the content has little to do with the results in section 3.

Line 345, The author need to state clearly what is validation set and test set because they mean different things. Meanwhile, if the conclusion was drawn based on fixed set of 8 test cases, it is highly likely that the conclusion won’t hold if test cases were shuffled with training data.

Author Response

We carrefully analyzed the requirements of the reviewer and we tried to gives pertinent answers that would contribute to the clarification of the text. We thank you for the observation!

Kind regards!

 

Author Response File: Author Response.docx

Reviewer 2 Report

  1. Just because forward propagation neural networks are popular, should not be a reason to use them. There is RNN, CNN, LSTM that are also popular and good at learning from examples. Please justify mathematically and correctly better as to why only forward propagation NN was chosen for this study. Alternatively, please consider further analysis with other types of NN and related results comparison demonstration.
  2.  Were other transfer functions such as renu etc. considered? Why or why not? Please justify. 
  3. In the results section please explain the comparison found clearly. For example, in the figure captions for figure 8 and 9, what was found? Please elaborate. Please consider to increase font size of the axis' labels.  
  4. Please consider reducing the discussions' section length.
  5. Please consider mentioning the key numerical findings in the conclusion section to support it as reported in the results section 

Author Response

We carrefully analyzed the requirements of the reviewer and we tried to give pertinent answers that would contribute to the clarification of the text. We thank you for the observations that contributed to the improvement of the work.

Kind regards!

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

line 226, like pointed earlier, training of neural network model should not rely on observing training error only as this will surely lead to overfitting or setting to a learning rate that was too low. In addition, validation is part of the training process to prevent overfitting. The author was using the validation to evaluate model performance instead. I understand that the authors may not have access to other software platform/programming except the commercial software. However, this could be a serious flaw in methodology that needs be overcome.

line 237, additional prediction? where is the prediction result?

line 311, best performing?

line 312 when saying very good correlation, the author is not specific as the standard of being good could be significantly different amongst people and domains.

line 320, another dataset never used in training and validation? Could you describe it in your experimental design section?

line325 table 3. the data format was inconsistent. what are the abbreviations?

line 331. more data often helps. However, it is not true that a larger number of hidden layers would also be helpful.

line 343, what is this new data?

Author Response

Thak you for your valuable proposal! Your suggestions improved our manuscrips. 

Thank you!

Nicoleta Anton 

Author Response File: Author Response.docx

Round 3

Reviewer 1 Report

Artificial neural networks (RNA)? Abbreviation should have been ANN, not to be confused with Ribonucleic acid  (abbreviated RNA).

There might be excess description of fundamental neural network concepts, which could be further condensed.

data format is inconsistent in Table 3.

Author Response

Dear reviewer,

Please consult the attached Response Letter regarding your comments.

Kind regards!

Nicoleta Anton

Author Response File: Author Response.docx

Back to TopTop