Development of Machine Learning for Asthmatic and Healthy Voluntary Cough Sounds: A Proof of Concept Study
Round 1
Reviewer 1 Report
This paper features the cough classification problem using machine learning techniques.
The off-the-shelf feature extraction of MFCC and CQCC, and the classifier of GMM-UBM are adopted. And the experiments and discussions are given in details.
Since all the technical development is based on the off-the-shelf techniques and there exist study on machine learning based cough analysis, the only contribution resides in the new application of cough classification and the data gathering.
Please do include more experiments with the state-of-the-art ML techniques instead of singly relying on the GMM-UBM.
Author Response
Response:
We thank the reviewer for the comment.
We have used GMM-UBM technique instead of the state-of-the-art machine learning techniques (such as Deep Neural Network, Convolutional Neural Network and Recurrent Neural Network) for two reasons in this report: Firstly, this is a proof of concept study to explore machine learning applied to paediatric cough sound classification. Second, GMM-UBM technique has been shown to perform reliably even with a smaller data samples [reference 39 in manuscript], while a richer dataset is required for the more recent state-of-the-art deep learning techniques.
Taking the reviewer’s comment, with more cough data being collected, we will explore other state-of-the-art machine learning approaches in future, and the results will be compared against other existing models. We look forward to reporting the results.
Reviewer 2 Report
It is an interesting paper on machine learning for asthmatic and healthy voluntary cough processing.
However the following comments need to be addressed before the paper can be processed any further:
1. Since the authors propose a GMM-UBM they should consider the state of the art in the field too. They should compare to papers with HMM-UBM which is able to provide better modeling and classification.
Please see papers such as "A novel holistic modeling approach for generalized sound recognition",
"Universal background modeling for acoustic surveillance of urban traffic", etc.
The data selection scheme could be very useful.
2. Then, seeing the trend in audio classification, i.e. from the DCASE competitions, a DNN classifier should be considered too.
3. The format of the journal is not respected.
4. It would be nice if the dataset was available in order to assess the difficulty of the task.
5. There are many naive mistakes, such as Table 2 is not complete.
6. The quality of the figures is poor. I suggest to use eps format.
7. DET curves could be easier to read than ROC ones. See "The DET curve in assessment of detection task performance"
Author Response
COMMENT 1. Since the authors propose a GMM-UBM they should consider the state of the art in the field too. They should compare to papers with HMM-UBM which is able to provide better modeling and classification. Please see papers such as "A novel holistic modeling approach for generalized sound recognition", "Universal background modeling for acoustic surveillance of urban traffic", etc. The data selection scheme could be very useful.
Response
We thank the reviewer for the comments.
There were previous studies on respiratory sound classification. One of which [1] used the vector quantization (VQ) method to classify respiratory sounds into those containing wheezes or normal respiratory sounds. Reported performance ranged from 60 to 70% using the VQ based classifier with audio features. Another study [2] used ANN to analyse breathing sounds (asthmatics during exacerbation, asthmatics in remission, and controls) and reported that ANN was able to deliver a classification accuracy of 43%. We believe that the low accuracy of ANN could be either attributed to the small number of subjects (50 subjects with asthma and 10 controls, ANN was trained using 72 vectors) or not having an optimized back propagation algorithm, activation function (given the fact that [2] was published in 1999). Compared with previous reports, the classification method GMM-UBM used in our study yielded a better classification accuracy result that exceeded 80%.
Taking the reviewer’s comment, other state-of-the-art machine learning approaches will be explored during our ongoing work with while a larger cough data are being collected.
§ [1] Bahoura, M., & Pelletier, C. (2003, May). New parameters for respiratory sound classification. In CCECE 2003-Canadian Conference on Electrical and Computer Engineering. Toward a Caring and Humane Technology (Cat. No. 03CH37436) (Vol. 3, pp. 1457-1460). IEEE.
§ [2] Rietveld, S., Oud, M., & Dooijes, E. H. (1999). Classification of asthmatic breath sounds: preliminary results of the classifying capacity of human examiners versus artificial neural networks. Computers and Biomedical Research, 32(5), 440-448.
COMMENT 2. Then, seeing the trend in audio classification, i.e. from the DCASE competitions, a DNN classifier should be considered too.
Response
We have used GMM-UBM technique instead of the state-of-the-art machine learning techniques (such as Deep Neural Network, Convolutional Neural Network and Recurrent Neural Network) in this report for two reasons: First, this is a proof of concept study to explore machine learning applied to paediatric cough sound classification. Second, GMM-UBM technique has been shown to perform reliably even with a smaller data samples [reference number 39 in the manuscript], while a richer dataset is required for the more recent state-of-the-art deep learning techniques.
Taking the reviewer’s comment, we agree that with more cough data being collected, other state-of-the-art machine learning approaches will be explored in future, and the results would be compared against other existing models.
COMMENT 3. The format of the journal is not respected
Response
Taking the reviewer’s comment, the format is modified to fit the journal’s requirements.
COMMENT 4. It would be nice if the dataset was available in order to assess the difficulty of the task.
Response
We regret the dataset cannot be made available to the public due to ethical constraints in Human Biomedical Research.
COMMENT 5. There are many naive mistakes, such as Table 2 is not complete.
Response
Taking the reviewer’s comment, we have updated the tables and figures in this resubmission.
COMMENT 6. The quality of the figures is poor. I suggest to use eps format.
Response
We agree with the reviewer’s comment. The quality has been improved and eps format will be available.
COMMENT 7. DET curves could be easier to read than ROC ones. See "The DET curve in assessment of detection task performance"
Response
ROC curves show the relationships between true positives and false positives as the decision threshold varies whereas DET curves plot miss rate (i.e., 1- true positives) verses false positives using the normal deviate scale. DET curves are widely used in the speaker recognition arenas while ROC curves are commonly used in medical journal arenas. We have used ROC curves in this report as medical practitioners and data mining specialists in general are more familiar with ROC curves compared to the former.
Reviewer 3 Report
Quality of figures and tables is quite poor. The definition of them must be improved.
Minor English changes may be required. A profesional proofreader revision could be useful.
Author Response
COMMENT 1. Quality of figures and tables is quite poor. The definition of them must be improved.
Response: The figures and tables have been updated and the quality has been improved. Figures in eps format will be available.
COMMENT 2. Minor English changes may be required. A professional proof reader revision could be useful.
Response: Taking the reviewer’s comment, a thorough proofreading was performed.
Round 2
Reviewer 2 Report
I would like to thank the reviewers for trying to address my comments.
Most of the changes are acceptable but still there is gap in the state of the art. The authors do not mention recent papers like “Classification of Sounds Indicative of Respira-tory Diseases” and the references within.
Also since your reply is "state-of-the-art machine learning approaches will be explored during our ongoing work with while a larger cough data are being collected" you have to at least mention those in the conclusions/discussion and describe their application. At the same time, the rationale behind not using them must be explained in the text.
These important changes will place this work better in the field and make it more consistent with the rest of the literature. Right now the papers included in the response date back to 2003 and 1999!
Author Response
We thank the reviewer and has made revisions.
We chose to work with GMM-UBM for a few reasons. Firstly, this is a proof of concept pilot study to explore machine learning in cough sound classification within a paediatric population. Most respiratory sounds investigated thus far in literature are breathing sounds or lung sounds rather than cough sounds. Secondly, GMM-UBM technique has proven its efficiency in respiratory sound classification [39,40,41,42], and found to perform reliably even with small data samples. While some recent studies show that deep learning approaches such as Convolutional Neural Networks are more efficient for respiratory sounds classification, notably lung sounds, these algorithms require a large dataset [43,44,45,46,47]. In [48], Bhattacharya reported that for a smaller dataset, UBM-GMM outperforms deep, recurrent neural net models. Given the size of our dataset, the pilot nature of the study in children, we opted for UBM-GMM. In future work, when our dataset grows larger, we will explore how deep learning approaches can further improve prediction accuracy.
We have included the response in the text (lines 366- 376)
We have included recent references.
39. Bahoura, M.; Pelletier, C.; Respiratory sounds classification using Gaussian mixture models. In Canadian Conference on Electrical and Computer Engineering 2004 (IEEE Cat. No. 04CH37513) (Vol. 3, pp. 1309-1312).
40. Mayorga, P.; Druzgalski, C.; Morelos, R. L.; Gonzalez, O. H.; Vidales, J.; Acoustics based assessment of respiratory diseases using GMM classification. In 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology (pp. 6312-6316).
41. Bahoura, M.; Pattern recognition methods applied to respiratory sounds classification into normal and wheeze classes. Computers in biology and medicine 2009, 39, 824-843.
42. Sen, I.; Saraclar, M.; Kahya, Y. P.; A comparison of SVM and GMM-based classifier configurations for diagnostic classification of pulmonary sounds. IEEE Transactions on Biomedical Engineering 2015, 62, 1768-1776.
43. Aykanat, M.; Kılıç, Ö.; Kurt, B.; Saryal, S.; Classification of lung sounds using convolutional neural networks. EURASIP Journal on Image and Video Processing 2017(1), 65.
44. Folland, R.; Hines, E.; Dutta, R.; Boilot, P.; Morgan, D.; Comparison of neural network predictors in the classification of tracheal–bronchial breath sounds by respiratory auscultation. Artificial intelligence in medicine 2004, 31, 211-220.
45. Güler, İ.; Polat, H.; Ergün, U.; Combining neural network and genetic algorithm for prediction of lung sounds. Journal of Medical Systems 2005, 29, 217-231.
46. Chen, C. H.; Huang, W. T.; Tan, T. H.; Chang, C. C.; Chang, Y. J.; Using k-nearest neighbor classification to diagnose abnormal lung sounds. Sensors 2015, 15(6), 13132-13158.
47. Ntalampiras, S.; Potamitis, I.; Classification of Sounds Indicative of Respiratory Diseases. In 2019 International Conference on Engineering Applications of Neural Networks (pp. 93-103). Springer, Cham.
48. Bhattacharya, G.; Alam, J.; Stafylakis, T.; Kenny, P.; Deep neural network based text-dependent speaker recognition 2016: Preliminary results. In Proc. Odyssey (pp. 9-15).
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
no further comments