Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This article targeted the suitability of data augmentation techniques for improving Parkinson’s disease detection with speech utterances. The research considered a binary classification between healthy subjects and subjects with Parkinson's symptoms. Convolutional neural networks (CNN) were considered. Researchers have used two datasets, both in Spanish. One dataset was dedicated to training and validation, whereas the other was reserved for testing only.

The paper reported an improvement beyond the baseline (no data augmentation) and a 3% increase with data augmentation. The baseline accuracy was 81%, and the accuracy with improved data, using augmentation, was 84%.

 

Comments:

1. Abstract must be a one paragraph stating all important components such as motivation, goal, methodology, objectives, and important outcomes. The authors should avoid itemized sections in the abstract unless the journal requires a different layout.
2. Are the two speech datasets public and reachable by researchers? I know they were cited in the references, but I am asking whether they are available to other researchers.
3. Speech frame size was set to 160 msec. For speech unit homogeneity, this is a very long frame size that mixes voiced, unvoiced, and silence units within a single frame; thus, it yields a false conclusion about speech features such as pitch value, its variation, and its presence.
4. Also, the given frame overlap (80 msec) is going to cause big jumps in processing the speech signal.
5. Is the silence portion between speech words and other units removed or processed as part of useful speech portions? I think this will affect system performance. On the other hand, did the researcher use speech endpoint detection in order to remove non-speech parts in the recordings?
6. “The system is just a binary classifier, which means that achieving good accuracy is an easy matter". What is the response of authors to this statement?
7. What is the relation between the four evaluation metrics listed in line 275? Ans, what are the outcomes from each one of them? Depending on the paper, the datasets are balanced with respect to all important factors.
8. CNN architecture is an old model, but more advanced models are available for classification purposes, especially ones with transformers and attention mechanisms. Why do the authors stick to the CNN model only?
9. It is good if the authors provide some information about the used CPU/GPU, CPU time required for both training and testing phases.
10. Is a kernel size of 3x3 the best for this research problem? Are there any comparisons/ investigations made?
11. Most of the references are slightly old. It is better to consider more recent references.

 

Author Response

We have included the responses to all reviewers' and editor's questions/comments in the attached file.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

At first glance, the manuscript appears highly relevant as it addresses the interesting and important topic of early detection of the Parkinson’s disease (PD) detection based on speech; however, its focus becomes unclear upon a full reading. A thorough revision addressing the following comments is required before the publication of the manuscript.

Major Comments:

1. The introduction provides a well-written overview of speech based early detection of PD, but its current structure may be misleading for readers, as the primary contribution of the study seems to be a comparative evaluation of different data driven, Artificial intelligence(AI)-based techniques rather than a focus on clinical implication in PD itself. The introduction would benefit from beginning with the machine-learning perspective at an earlier stage and then introducing the details of PD later. Such a restructuring would better align the introduction with the study’s objectives.
2. The data augmentation techniques have their own inherent limitations, as they may fail to adequately capture clinical parameters of a disease such as speech characteristics and cannot compensate for small datasets regarding limited demographic details or pathological variability. In recent years machine learning techniques such as transfer learning have demonstrated strong performance in clinical setting with small datasets using large-scale pretraining and often surpassing models trained from scratch. In this context, it would be beneficial for the authors to clarify the rationale for prioritizing independent dataset evaluation over the use of established machine learning approaches that are specifically designed to address small-sample scenarios.
3. The manuscript does not sufficiently discuss the clinical implications of the reported findings. It remains unclear how the observed results should influence current perspectives on speech-based PD screening in clinical settings. A discussion about how these findings could alter clinical decision-making would significantly enhance the practical relevance of the study. Furthermore, the manuscript would also benefit from a discussion of ethical considerations, including risks related to misdiagnosis, false reassurance, and data privacy.

Minor Comments

1. The caption of Figure 2 should be revised to more explicitly describe the specific effects of the data augmentation techniques.
2. To strengthen the manuscript, the authors should include recent research on different AI based augmentation methods for pathological speech by citing the relevant and up-to-date studies.

Author Response

We have included responses to all questions/comments from the reviewers and the editor in the attached file.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Orozco-Arroyave et al present research on the significance of data augmentation techniques to improve Parkinson’s disease (PD) detection with speech recordings. I have the following comments regarding this manuscript:

1. In the introduction a broader view on PD pathogenesis, background, prodromal symptoms would be beneficial
2. It would be valuable to provide a perspective on the possible relevance data augmentation techniques in the context of speech recording in diseases closely related to Parkinson's disease e.g. Progressive Supranuclear Palsy-Parkinsonisms Predominant.
3. Authors could provide a wider perspective on the speech analysis in the context of its possible links to endoscopic evaluations - Ref. Oropharyngeal dysphagia in Parkinson's disease: potential utility of speech acoustic analysis in detection and evaluation of swallowing impairment progression. Neurologia i neurochirurgia polska, 59(5), 478–486. https://doi.org/10.5603/pjnns.105102
4. The suitability of implementing different data augmentation should be analyzed in the context of PD subtypes - postural instability/tremor-dominant etc.
5. The limitation section should be more stressed and presented in a critical manner.

Author Response

We have included responses to all questions/comments from the reviewers and the editor in the attached file.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

Authors implemented some of the correction, however are encouraged to implement the third suggestion from the previous round of review.