Combining Supervised and Unsupervised Fuzzy Learning Algorithms for Robust Diabetes Diagnosis

Round 1

Reviewer 1 Report

The authors propose a new hierarchical combination of classifiers that is supposed to performed better than current classifiers. 

Although the idea is intereseting and the classifier algorithm itself may valuable, the article presents notable methodological flaws:

- As the purpose of the proposed classifier algorithm is to overcome the limatations of current classifiers in preventing overfitting, a validation set (on top of a training set and test set) is necessary to assess the actual model performance. Notably, the database in use is likely to be too small to provide a validation set so that new data should be seeked.

- Authors conclude that the proposed algorithm performs better than existing algorithms in different accuracy metrics but do not provide any statistical test comparing performances that justifies these statements.

- It is not clear what is the event that the classifier predicts. It seems reasonable to guess it is having or not having diabetes but this should be clearly declared in the 3.2 section

- As suggested in the "instruction for authors" page of the journal, research manuscripts should be organised according to the following sections: Introduction, Materials and Methods, Results, Discussion, Conclusions (optional).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

1.   The Authors proposed a hierarchical combination of fuzzy C‐means clustering component and fuzzy Max‐Min neural network supervised learner for robust diabetes diagnosis.

2.   All mathematical formulas, equations and symbols used in this paper are correctly defined.

3.   The successive steps of proposed methods and the way of solution are correctly described in the form of Algorithms 1 and 2 and illustrated in Figure 2 with sufficient details.

4.   The numerical calculations were performed on the open Pina Indian Diabetes (PID) database containing 768 female diabetic patients’ information. The experiment results are properly depicted in Tables 1-3 with correct comments.

5.   The proposed fuzzy hierarchical combination FCM‐FMM method is competitive in comparison with various supervised algorithms, reported in the literature, in the highly noisy problem domain like Diabetes diagnosis using the same PID database and the same measures (sensitivity, specificity, accuracy, F1 score). In further studies it is necessary to achieve a better accuracy, more than 80 %.

6.   Remarks:

·     The abbreviation FCM should be deciphered in line but not at the and of this article in lines 194-195.

·     The abbreviation TP, TN, FP, FN and words Precision and Sensitivity are written in the text in normal font, in tables in italic font.

·     The parameter ‘Precision’ in Table 2 for F1 Score is not defined. Should not the word ‘Accuracy’ be there?

·     Editing errors in the words: ‘unsupertvised’ in line 190, performance in line 206 and hierarchical in line 210.

·     All mathematical symbols and indexes in text and algorithms should be written in italic font, e.g., N, p, y, w.

·     The word ‘where’ in Algorithm 2 Step 1 not italic font.

·     The notation of vectors in the description of Algorithms 1 and 2 should differ from their components, e.g., x, y, w. With such notation, it is very difficult to analyze the correctness of the presented formulas. They should be written, for example, in bold fonts.

·     In Figure 2 in the second block is the symbol #. Is it correct? Shouldn't the parameters from Algorithm 1 from step 1 be given there?

·     Are the outputs (No, Yes) from the conditional block (Data available?) marked correctly in Figure 2?

·     The Authors did not specify what kind of activation function was used in their back propagation neural network.

7.   The Authors presented the problem and proposed its solution together with experimental results which can be published after correcting these indicated editing errors.

Comments for author File: Comments.docx

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Overall, the quality of the paper has been strongly improved by the authors.

A brief discussion on why the decision not to perform an internal validation was taken may be added to the limitations paragraph.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The Authors proposed a new hierarchical combination of fuzzy C‐means clustering component and fuzzy Max‐Min neural network supervised learner for robust diabetes diagnosis that is supposed to performed better than current classifiers.

The Authors revised the manuscript very carefully according to my remarks, modified the mathematical expressions to make the description of the article more understandable and made a lot of the detailed changes listed in the answers to Reviewers.

In order to verify that the proposed algorithm performs better than existing algorithms in different accuracy metrics, they used additional one-way ANOVA (Analysis of Variance) test and subsequent Tukey test. The results of additional simulation tests confirmed that the proposed algorithm performs better than existing algorithms in different accuracy metrics.

The conclusions of this paper are also reorganized based on the additional calculations.

The Authors consider the problem and proposed its solution which is relevant and interesting for publication in this journal.

Round 3

Reviewer 1 Report

I think the article has been strongly improved.