Deep Learning and Recurrence Information Analysis for the Automatic Detection of Obstructive Sleep Apnea

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article presents the following concerns: 

• Although the paper uses well-known public databases (Apnea-ECG, MIT-BIH, and UCD-DB), it would be useful to discuss in more depth how each database's demographic and technical characteristics affect the generalization of the model. What biases could these differences introduce into the final model?

• Although the CNN-based model outperforms traditional ML approaches, the work could benefit from a more detailed description of how traditional models were optimized (e.g., selected hyperparameters and feature selection criteria).

• The erroneous cases (false positives and false negatives) should be discussed in depth. Analyzing these errors could reveal interesting patterns or biases that could be improved in future iterations of the model.

• Although the limitation of direct comparison due to methodological differences is acknowledged, it would be useful to include a summary table explaining how this approach adds value over other works, considering key metrics such as sensitivity, specificity, and accuracy.

• The introduction does a good job of contextualizing the importance of obstructive sleep apnea and the need for alternatives to PSG. However, the transition between existing methods and the justification for choosing CNNs and recurrence analysis could be strengthened.

• The section on parameter selection, such as τ and m, could benefit from a deeper discussion on how the selected values ​​affect the results and their generalization to other contexts.

• A short paragraph on specific possible future steps, such as improvements in model generalization or more robust data acquisition, would be valuable.

• The figure caption needs to be improved.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This paper proposed a deep learning and recurrence information analysis for the automatic detection of obstructive sleep apnea, and the effectiveness of the method has been proved based on publicly-available databases. In my opinion, the work is very meaningful, and it can be published after revision. Some suggestions are given as follows:

1. Besides Fig. 1, it is suggested that the author can add a figure of the deep learning process, such as most deep learning works.

2. For deep learning, a core issue is that how to determine the hyper-parameters and learning rate, and how did the author do? Some descriptions can be added.

3. Did the author do some improvements to improve the robustness of the method? Because in practice, there would be many unknown disturbances beyond the databases, and how to ensure the robustness of the method?

4. CNN and other neural networks, such as LSTM, can also process the timing signal, thus, did the author consider process the ECG recording directly?

5. An important issue of deep learning is model degradation after a long period, the proposed method also faces this problem. Some other works discusses the solution, (such as re-training: Surface roughness prediction of large shaft grinding via attentional CNN-LSTM fusing multiple process signals). It is suggested that the author can consider this problem in Discussion part, and the existing method in above paper could be cited.

6. The accuracy is about 75%, which seems can be further improved, thus, it is encouraged that the author can give an outlook of future work to tell us how to further improve the accuracy in theory.

Author Response

Please see the attachment

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript has been improved greatly.

Reviewer 2 Report

Comments and Suggestions for Authors

No comments.