Abnormal Heart Sound Classification and Model Interpretability: A Transfer Learning Approach with Deep Learning

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

I have read your manuscript with great interest. However, as it is now, it somehow confuses me and leaves me with a lot of questions. Generally, try to think as a reader looking at it for the first time - everything should be cristal clear. Here are some specific comments:

-) Introduction: Why is your introduction expanded by a "literature survey"? I am all for giving an overview, but this seems extensive. You could, however, re-structure your article and provide results in two stages: 1) A narrative review of the state of the art (but then, also other artificial heart sound detection / interpretation that has been published must play a bigger role to really paint a picture of what's already out there), and 2) Your investigative results. Please then adapt the Methods accordingly (even a narrative review should have some kind of pico etc.).

-) Methods: It is entirely unclear to me what exactly you did here. Where do the patients come from? Did you develop these algorithms or is this a feasibility study of an already-existing one? Etc... A reader must be guided through your manuscript without asking themselves these questions.

Comments on the Quality of English Language

No major issues, just try to reduce complexicity of sentences.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

**Comments on the Paper: "Abnormal Heart Sound Classification and Model Interpretability: A Transfer Learning Approach with Deep Learning"**

The paper presents a comprehensive study on the classification of abnormal heart sounds using a transfer learning approach with deep learning techniques. The authors address the challenges of accurately detecting cardiovascular disease through heart sound analysis, highlighting the limitations of traditional auscultation methods due to noise and dataset scarcity. They propose a method that combines phonocardiogram (PCG) and electrocardiogram (ECG) data, fine-tunes pre-trained deep convolutional neural network (CNN) models, and provides an analysis of model interpretability.

**Strengths:**

1. **Relevance:** The paper addresses an important problem in the medical field, leveraging machine learning to enhance heart sound abnormality detection. This is relevant given the increasing interest in using AI for healthcare applications.

2. **Experimental Approach:** The authors meticulously describe their methodology, from data preprocessing to model training and evaluation. The use of various image representations (spectrograms, mel-spectrograms, scalograms) and pre-trained models (ResNet, VGG, inceptionv3) provides a comprehensive analysis.

3. **Performance Improvement:** The results indicate a significant improvement in accuracy compared to previous methods. Achieving 91.25% accuracy on the training dataset is notable and demonstrates the efficacy of the proposed approach.

4. **Interpretability:** The inclusion of model interpretability analysis using methods like guided back-propagation, Grad-CAM, and LIME adds value to the paper. This provides insights into which features are driving the model's decisions, enhancing its clinical relevance.

**Areas for Improvement:**

1. **Dataset Description:** While the paper mentions using the PhysioNet Computing in Cardiology Challenge 2016 dataset, a more detailed description of the dataset's characteristics and size would help readers better understand the experimental setup.

2. **Discussion of Challenges:** The paper briefly mentions the challenges posed by noise and limited datasets. Further elaboration on these challenges and their impact on model performance would enhance the paper's credibility.

3. **Comparison with Literature:** The authors mention that their method outperforms a previous CNN method by Rong et al. It would be helpful to provide a more detailed comparison with other related works in the literature, discussing their strengths and weaknesses.

4. **Visualization of Results:** The paper could benefit from the inclusion of visual aids such as figures and tables to showcase the quantitative results in a more organized and easily digestible manner. The captions of all graphics and tables need to be addressed thoughtfully.

5. **Ethical Considerations:** Given that medical AI applications can have ethical implications, discussing potential limitations, biases, and ethical considerations related to the use of AI in medical diagnosis would provide a more holistic perspective.

6. **Future Directions:** Including a section discussing potential future directions for research based on the findings of this study could provide insights for researchers interested in building upon this work.

**Overall Impression:**

The paper presents a well-structured study on abnormal heart sound classification, utilizing transfer learning and deep learning techniques. It offers valuable insights into the methodology, experimental results, and model interpretability. By addressing the limitations and challenges of previous methods, the paper contributes to the field of medical AI. To enhance its impact, the paper could further improve its dataset description, discuss the limitations in-depth, and contextualize the results within the broader literature.

Comments on the Quality of English Language

The quality of English language in the paper is generally good, with clear and concise writing. However, there are a few areas where improvements could enhance the readability and clarity of the content:

1. **Sentence Structure:** Some sentences are quite long and complex, which could make the content more challenging to follow. Consider breaking down complex ideas into shorter sentences for better comprehension.

2. **Clarity of Expression:** While the paper provides a detailed explanation of the methodology, there are instances where certain explanations could be further clarified to ensure that the reader fully understands the process.

3. **Consistency in Terminology:** Ensure consistent usage of terminology throughout the paper. For example, consistently using "heart sound" instead of alternating between "heart sound" and "heart signals."

4. **Transition Sentences:** Include transitional sentences or phrases to guide the reader smoothly from one section or concept to another. This can improve the flow of the paper.

5. **Visual Aids:** Consider incorporating figures or tables to visually present complex results or processes. Visual aids can help the reader grasp the content more easily.

6. **Dataset Details:** Provide more detailed information about the dataset used, including its source, size, and characteristics. This will help the reader understand the context of the study better.

7. **Ethical Considerations:** Given the medical nature of the study, it would be beneficial to address potential ethical considerations related to patient data, model interpretability, and real-world deployment.

8. **Proofreading:** While the paper is generally well-written, there are instances of minor grammatical errors or awkward phrasings. A thorough proofreading could help eliminate these issues.

Overall, the paper effectively communicates its methodology and results, but with some minor improvements in sentence structure, clarity, and terminology consistency, it could further enhance its readability and impact.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for the opportunity to review the presented article.

Overall an interesting article. Please read the article carefully as there are punctuation errors. Discussions should be expanded to include available studies comparing physical examination results with imaging tests such as echocardiography.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The article submitted for consideration has an undoubted relevance and novelty, its methodology is clearly described, the results obtained are convincingly presented and interpreted in 7 tables and 20 figures.

The relevance lies in the fact that for quick and high-quality diagnosis of the most common pathology in humans - cardiovascular diseases - it is necessary to improve screening methods based on modern achievements of advanced science, which include the use of artificial intelligence and, in particular, one of the ways to implement it - neural learning networks. To date, a number of training methods and network architectures have been developed, but at the same time, the advantages of one or another approach for use in cardiology are not obvious. Therefore, further research is needed in this direction.

The authors explored the capabilities of machine learning models in classifying abnormal heart sounds using PCG and ECG signals from the CinC 2016 training dataset, presented characteristics of their specificity, sensitivity, accuracy, and also showed that using engineering functions to split PCG into four bands and combine this with ECG allows you to achieve the highest accuracy of the classifier during its training (has extended the performance of machine-learning models in classifying abnormal heart sounds using PCG and ECG signals from the CinC 2016 training-a dataset).

The conclusions are substantiated and the goal set as a whole is disclosed, the direction of further research in this area is outlined.

Strengths of the study: new data based on the analysis of the performance of the studied models regarding the use of machine learning models in cardiology, given the paucity of studies in this branch of science; qualitative representation of the controlled backpropagation method in the form of numerous Grad-CAM and LIME images for the optimal inceptionv3 classifier, established on the basis of own data; high potential for practical use of the results, given their use for screening the most common pathology in humans.

Weaknesses of the study: it is not described how the data were collected and labeled to exclude the subjective assessment of the expert, in particular, how many experts participated in the implementation of this research step for each FCG and ECG data set.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Dear authors,

Thank you for making the manuscript more clear and readable for someone not being an expert on the topic (such as the broader scientific community).

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you very much for the opportunity to review the article. The article is well written. In my opinion, it will interest readers and inspire them. Thank you