Machine Learning for Automated Classification of Abnormal Lung Sounds Obtained from Public Databases: A Systematic Review

Round 1

Reviewer 1 Report

 

The systematic review provides a comprehensive overview of the use of machine learning (ML) models for classifying abnormal lung sounds derived from public databases. The identified background underscores the significance of pulmonary auscultation in diagnosing lung abnormalities during initial physical assessments, while highlighting the operator-dependency and inter-observer variability challenges associated with traditional auscultation. The objective of this systematic review is well-defined, aiming to compare the characteristics, diagnostic accuracy, concerns, and data sources related to ML-based classification of abnormal lung sounds using recordings from public databases.

However, the methodology is not adequately described, encompassing a comprehensive literature search across multiple reputable databases and subsequent screening, review, and data extraction processes. The use of a modified QUADAS-2 tool for quality assessment is appropriately described. The results section briefly presents key findings, including the prevalence of artificial neural networks (ANN) and support vector machines (SVM) as the most commonly employed ML classifiers. The provided accuracy ranges for different classification tasks highlight the variability in model performance across studies, thereby offering a realistic perspective on the achieved outcomes. The identification of seventeen public databases provides valuable insights into available data sources, with a notable emphasis on the ICBHI 2017 database. However, the discussion should raise critical points regarding the study. 

Specific Comments: 

Abstract should be 200 words long.

Please provide an updated PRISMA checklist.

Do you mention a protocol for your work? Have you drafted a protocol before starting data extraction? If so, please specify it in the text and provide the link to where it is accessible.

Please also specify how many people carried out the bibliographic research and whether any information specialists were consulted. Selection criteria

 

Please specify how many authors carried out the selection of studies and whether this was done in duplicate. If more than one author was involved, please specify how concordances were assessed and how discrepancies between them were resolved.

Please specify how many authors performed the data extraction and whether it was performed in duplicate. If more than one author was involved, please specify how concordances were assessed and how discrepancies between them were resolved.

Please specify how many authors performed the quality assessment of the studies according to your tool and whether this was done in duplicate. If more than one author performed the assessment, please specify how concordances were assessed and how discrepancies between them were resolved.

Figure 3 is not clear. Caption should explain both plots clearly.

Result section is brief and not highlighting key findings properly. The results report findings without interpretation, while the discussion analyzes, criticizes, and discusses them.

Limitation and future directions are missing.

You mentioned publication bias but the tool for assessment is not objective.

 

I strongly suggest adding a PICOS table (e.g, see Table 1 in https://www.frontiersin.org/articles/10.3389/fmed.2023.1109411/full) to complement section 4.2.

Author Response

Please see the attached cover letter for the responses. 

In response to the comment requesting an updated PRISMA 2020 Checklist, we attach the corresponding PRISMA table. The rest of the responses are in the attached cover letter. 

 

Author Response File: Author Response.pdf

Reviewer 2 Report

This study proposes a systematic review of abnormal lung sound classification with ML models and recordings derived from public databases to compare differences in characteristics, diagnostic accuracy, concerns, and data sources.

 

In reviewing the studies, the review results should be explained in detail.
I have some suggestions regarding this study:
1. Please provide more detailed methods used by the literature in the result table. It can be the feature extraction method used, and so on.
2. Please provide the results achieved by the literature (the maximum scores of accuracy).
3. Please provide the strengths of each reviewed study (literature).
4. Please provide the limitations of each piece of literature reviewed in the study.
5. Please strengthen the clinical relevance.
6. Please discuss the possible future challenges and work in the area of abnormal lung sound classification and detection based on the results achieved from the review.

Author Response

Please see attached cover letter with the responses to each suggestion. 

Author Response File: Author Response.pdf

Reviewer 3 Report

1.   The review provides a comprehensive update on using contemporary and DL models and highlights the advances on automatic lung sound classification focusing on the introduction of large public databases. A good approach. Also large public data sources in recent years led to an increasing number of studies to share their lung sound audio samples, ideally facilitating comparison between models is well shown.

2. Your work highlights the models identified in systematic review with the best accuracy, sensitivity, and specificity performance metrics. A table/graph on accuracy improving and sensitivity is required and it will add value to the work.

3. Machine learning (ML) and Deep learning (DL) techniques are good for identification and classification of normal and abnormal lung sounds and measured with degrees of diagnostic certainty depending on the experience level and skill set. The inability to identify and accurately classify lung sounds could significantly impact the delay in diagnosis and downstream management and as such the approach will be effective as they employed a multi-layer perception neural network employing a back propagation training algorithm to predict normal or abnormal lung sounds. How do you justify this fact.

4.   Meta-analysis, given the heterogeneity in performance reporting and data sources is under suspect for large portion of the older databases, what is your suggestion to address this issue.

5. How will you comply with the concept with further advancements in computational process need to be justified as these techniques have the potential to provide better-personalized precision medicine and accurate assessment of respiratory conditions, aiding in diagnoses, monitoring, and treatment?

Author Response

Please see the responses in the attached cover letter. 

Author Response File: Author Response.pdf

Reviewer 4 Report

This paper performed a systematic review of abnormal lung sound classification with ML models and recordings

derived from public databases to compare differences in characteristics, diagnostic accuracy, concerns, and data sources.

I have following comments:

1.The sentence "Pulmonary auscultation is essential to identifying abnormal lung sounds in at initial physical assessment."

should be

"Pulmonary auscultation is essential to identify abnormal lung sounds in initial physical assessment."

2. In sections 1.2.2~1.2.4, i suggest that the authors should give a comparison of each method, for example, advantage and drawback.

Author Response

Please see the responses in the attached cover letter. 

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

I would like to thanks for their nice work. Authors have addressed all of previous comments. I have no further comments. 

Reviewer 2 Report

All the comments and suggestions have been addressed

Reviewer 4 Report

No comments

No comments