Identification of Parkinson’s Disease from Native Italian People: Machine Learning Voice Analysis

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study evaluates classical machine-learning classifiers for distinguishing Parkinson's disease (PD) speech from healthy controls using the publicly available Italian Parkinson's Voice and Speech (IPVS) dataset.

Major concerns:

• Validation strategy is internally inconsistent: the manuscript variously claims a 75 / 25 split, GridSearchCV k-fold cross-validation, and LOSOCV. Only one of these can have produced the headline numbers, and the reader cannot tell which. This is the single most important issue to resolve.
• Severe age confounding is present in the design (Young HC 19–29 vs PD 40–80) and is not addressed. A classifier could be reading age, not PD.
• Several numerical inconsistencies appear between abstract, tables, and figures (an F1-score that is not the harmonic mean of the reported precision and recall; varying confusion-matrix totals; MCC discrepancies). These need to be reconciled.
• No code repository, random seeds or hyperparameter grids are provided. The 'reproducible pipeline' claim is currently not backed by reproducibility infrastructure.

 Overall, the underlying study is commendable, and the GTCC-vs-MFCC finding is worth publishing  but the validation procedure, age-confound, and numerical inconsistencies must be addressed before this manuscript can carry diagnostic-utility claims. 

Comments on the Quality of English Language

NA

Author Response

Dear Editor,

We are writing to you regarding our paper "Identification of Parkinson’s disease from native Italian people: Machine learning voice analysis", which has been submitted to the Journal of BioMed for publication.

We would like to thank you for your helpful review comments that have aided us in improving our manuscript. Our point-to-point reply to the reviewers’ comments and suggestions is below.

Reviewer# 1

We thank Reviewer for the constructive feedback and thoughtful suggestions. Below we address each comment in detail and indicate where changes have been made in the manuscript.

Comments and Suggestions for Authors

This study evaluates classical machine-learning classifiers for distinguishing Parkinson's disease (PD) speech from healthy controls using the publicly available Italian Parkinson's Voice and Speech (IPVS) dataset.

Major comments:

• Validation strategy is internally inconsistent: the manuscript variously claims a 75 / 25 split, GridSearchCV k-fold cross-validation, and LOSOCV. Only one of these can have produced the headline numbers, and the reader cannot tell which. This is the single most important issue to resolve.

Response:

We thank the reviewer for highlighting this important inconsistency. We clarify that all reported results in this study were obtained using a subject-wise 75/25 train-test split, ensuring that all samples from each participant were contained exclusively in either the training or test set.

Hyperparameter optimization was performed only on the training set using GridSearchCV with 10-fold cross-validation. The best-performing model configuration was selected based on the average cross-validation score across training folds, with no access to test data during tuning. GridSearchCV with 10-fold cross-validation was used solely for hyperparameter optimization within the training set. For each classifier, hyperparameter combinations were evaluated using 10-fold cross-validation applied only to the training data. The optimal hyperparameter configuration was selected based on the average cross-validation performance across the training folds. No information from the independent test set was used during model selection, parameter tuning, or preprocessing parameter estimation.

Final performance metrics Accuracy, Precision, Recall, F1-score, ROC-AUC, and MCC were computed exclusively on the independent unseen test set.

References for LOSOCV were included in error and have now been removed. The Methods section has been fully revised to reflect a single, consistent evaluation pipeline. The methodology section has been thoroughly revised. Please see pages 5 and 6, Lines 197-219 and page 6 lines 230-248

• Severe age confounding is present in the design (Young HC 19–29 vs PD 40–80) and is not addressed. A classifier could be reading age, not PD.

Response:

We acknowledge that the IPVS dataset includes an inherent age imbalance between young healthy controls and PD participants, which may introduce potential confounding effects.

In our study, raw audio recordings were obtained across three groups: Young HC, Adult HC, and PD. Acoustic, MFCC, and GTCC features were extracted from all recordings using the same preprocessing pipeline. During model development, all healthy control data (including both young and adult participants) were consolidated into a single HC class, while PD recordings formed the PD class. No age-related or demographic variables were included in the feature set or model training process. Please See page 5, lines 170-174

Therefore, the classifiers were trained exclusively on acoustic feature representations without explicit access to age information. However, we agree that residual age-related speech characteristics may still indirectly influence acoustic features and thus cannot be fully excluded as a confounding factor.

To address this limitation, we have added an explicit discussion in the revised manuscript, clarifying that model performance may partially reflect age-related variability in speech and that results should be interpreted as dataset-specific discrimination rather than age-independent clinical diagnosis. We also highlight the need for future studies using age-matched cohorts and controlled experimental designs. Please see page 17 and line number 545-551.

• Several numerical inconsistencies appear between abstract, tables, and figures (an F1-score that is not the harmonic mean of the reported precision and recall; varying confusion-matrix totals; MCC discrepancies). These need to be reconciled.

Response:

We thank the reviewer for carefully identifying these inconsistencies in the reported performance metrics. We conducted a complete audit of all evaluation results across the manuscript. During this review, we identified transcription and rounding inconsistencies between the Abstract, Results section, tables, and figures.

To resolve this issue, all performance metrics were recomputed directly from the final model prediction outputs. This ensured full consistency between precision, recall, F1-score, confusion matrices, and Matthews Correlation Coefficient (MCC). We update results section with tables 3 to 5, figures 2 to10 and corresponding textual descriptions have been updated accordingly. Please go with page 8 lines 295-307, page 9 lines 312-323 and lines 328-332.  And also, please see page 8-14 results section for updated figures and tables.

• No code repository, random seeds or hyperparameter grids are provided. The 'reproducible pipeline' claim is currently not backed by reproducibility infrastructure.

Response:

We thank the reviewer for highlighting this important issue regarding reproducibility. We agree that reproducibility requires clear reporting of implementation details and experimental settings.

To address this concern, the manuscript has been revised to include comprehensive details necessary for replication. Specifically, we now report the random seed values used throughout the experimental pipeline to ensure consistent data splitting and model initialization. In addition, complete hyperparameter search spaces for all machine-learning classifiers are provided in the revised manuscript in Table 1, page 6.

We have also expanded the Methods section to include software environment specifications, including Python version and key library versions, as well as detailed parameter settings used for MFCC and GTCC feature extraction to ensure transparency of the preprocessing pipeline. Please go with pages 5 and 6, lines 158-167 and page 6, lines 177-180.

Regarding code availability, we clarify that the full implementation code is not publicly released due to project constraints; however, it can be made available from the corresponding author upon reasonable request under a data-use agreement. Please see page 19, Code Availability Statement. Line 642-644.

Reviewer 2 Report

Comments and Suggestions for Authors

This paper presents a machine learning study for detecting PD from voice recordings using the IPVS dataset. The authors extract acoustic, MFCC, and GTCC features and evaluate six traditional classifiers, achieving the best performance of 94.68% accuracy with SVM on combined acoustic + GTCC features under subject-independent validation.

While I find the topic highly relevant to the research community, I believe the novelty of this work is incremental rather than groundbreaking when compared to the current state-of-the-art (SOTA) on the IPVS dataset.

Most existing studies already employ similar acoustic features (including MFCCs), perturbation/prosodic/temporal features, and conventional machine learning classifiers (SVM, RF, XGB, etc.), frequently reporting high classification accuracies in the range of 91–99%. Several earlier works have even achieved performance exceeding 97–98% in certain configurations *(see the list of similar papers at the end of my comments).

The present manuscript offers a useful incremental contribution, particularly by:
1. Highlighting the value of Gammatone Cepstral Coefficients (GTCCs) in PD voice analysis (a relatively underexplored feature in this dataset);
2. Providing a clean, reproducible benchmark using traditional ML models with systematic feature fusion;
3. Reinforcing the potential of speech as a digital biomarker specifically on the Italian cohort.

However, the paper lacks major methodological breakthroughs. There is no new architecture proposed, no cross-lingual or cross-dataset evaluation in the main results (only mentioned as future work), no longitudinal analysis, and no truly novel feature engineering beyond the inclusion of GTCCs. Furthermore, the study is limited to binary classification (PD vs. HC) on a relatively small, publicly available dataset, an approach that is now quite common in this field.

Recommendation to the authors:
I strongly suggest that the authors clearly articulate a more focused and novel research question. Their current solid results should be positioned as a baseline or comparative benchmark against the SOTA. Relevant recent reviews that could serve as strong comparators include: https://doi.org/10.3390/app14177873, https://doi.org/10.3390/inventions10040048

Additionally, the authors should cite and discuss recent systematic reviews on machine learning and deep learning approaches using speech as a biomarker for PD detection to better contextualize their contribution.

Minor Comments:
1. Please substantially revise the abstract. It should be written as one single, coherent paragraph. Clearly articulate the topic, the specific gap addressed, and the main goal of the study. I recommend following the Golden Circle approach (Why → How → What).
2. There is no clear RQ/Hypothesis in the Introduction.
3. Title: Remove the final dot at the end of the title.

*List of similar papers:
1. Speech signals-based Parkinson’s disease diagnosis using hybrid autoencoder-LSTM models -Ayşe Nur Tekindor, Eda Akman Aydın (2025)
2. Interpretable Early Detection of Parkinson’s Disease Through Speech Analysis - L. Simone, M. G. Camporeale, V. M. Rubino, V. Gervasi, G. Dimauro (2025)
3. A Triplet Multimodel Transfer Learning Network for Speech Disorder Screening of Parkinson’s Disease - Aite Zhao, Nana Wang, Xuesen Niu, Huimin Wu (2024)
4. Leveraging Deep Learning for Fine-Grained Categorization of Parkinson’s Disease Progression Levels through Analysis of Vocal Acoustic Patterns - Hadi Sedigh Malekroodi, Nuwan Madusanka, Byeong-il Lee, Myunggi Yi (2024)
5. Exploring robust computer-aided diagnosis of Parkinson’s disease based on various voice signals - Jiu-Cheng Xie, Yanyan Gan, Ping Liang, Rushi Lan, Hongbing Gao (2022)

Comments on the Quality of English Language

It should be revised.

Author Response

Reviewer# 2

We thank Reviewer for suggestions, careful evaluation and feedback. We have worked to address all concerns.

Comments and Suggestions for Authors

This paper presents a machine learning study for detecting PD from voice recordings using the IPVS dataset. The authors extract acoustic, MFCC, and GTCC features and evaluate six traditional classifiers, achieving the best performance of 94.68% accuracy with SVM on combined acoustic + GTCC features under subject-independent validation.

While I find the topic highly relevant to the research community, I believe the novelty of this work is incremental rather than groundbreaking when compared to the current state-of-the-art (SOTA) on the IPVS dataset. 

Most existing studies already employ similar acoustic features (including MFCCs), perturbation/prosodic/temporal features, and conventional machine learning classifiers (SVM, RF, XGB, etc.), frequently reporting high classification accuracies in the range of 91–99%. Several earlier works have even achieved performance exceeding 97–98% in certain configurations *(see the list of similar papers at the end of my comments).

The present manuscript offers a useful incremental contribution, particularly by:

1. Highlighting the value of Gammatone Cepstral Coefficients (GTCCs) in PD voice analysis (a relatively underexplored feature in this dataset);
2. Providing a clean, reproducible benchmark using traditional ML models with systematic feature fusion;
3. Reinforcing the potential of speech as a digital biomarker specifically on the Italian cohort.

However, the paper lacks major methodological breakthroughs. There is no new architecture proposed, no cross-lingual or cross-dataset evaluation in the main results (only mentioned as future work), no longitudinal analysis, and no truly novel feature engineering beyond the inclusion of GTCCs. Furthermore, the study is limited to binary classification (PD vs. HC) on a relatively small, publicly available dataset, an approach that is now quite common in this field.

Recommendation to the authors:
I strongly suggest that the authors clearly articulate a more focused and novel research question. Their current solid results should be positioned as a baseline or comparative benchmark against the SOTA. Relevant recent reviews that could serve as strong comparators include: https://doi.org/10.3390/app14177873, https://doi.org/10.3390/inventions10040048

Additionally, the authors should cite and discuss recent systematic reviews on machine learning and deep learning approaches using speech as a biomarker for PD detection to better contextualize their contribution.

Response:

We thank the reviewer for this valuable feedback and agree that the primary contribution of this work is not the development of a novel classification architecture. Accordingly, we have revised the manuscript to position the study as a systematic comparative benchmark of acoustic, MFCC, GTCC, and fused feature representations using traditional machine-learning classifiers on the IPVS dataset.

To better clarify the study's scope and contribution, we have:

Added research questions in the Introduction to define the objectives of the study. Please see page 2-3 line 86-98.

Revised the Introduction and Discussion sections to emphasize the benchmark nature of the work and the relatively underexplored role of GTCC features within the IPVS dataset.

Revised the Abstract and Conclusion to avoid overstating novelty and to better contextualize the findings within the existing body of literature. Please go with introduction page 2-3, and Strengths and implications subsection in page 16. Alos we update Limitations and future directions subsection in pages 16-17, and Conclusion section in pages 17-18.

We believe these revisions more accurately position the study as a reproducible comparative evaluation that complements existing research while providing insight into the contribution of GTCC-based feature representations for PD speech analysis.

Minor Comments:
1. Please substantially revise the abstract. It should be written as one single, coherent paragraph. Clearly articulate the topic, the specific gap addressed, and the main goal of the study. I recommend follResponse

 Response:

We revised abstract as per journal guidelines. Please see page 1. The abstract has been completely revised and reformatted in accordance with the journal guidelines. The revised version clearly states the Background/ Objectives, Methods, Results, and Conclusions.

2. There is no clear RQ/Hypothesis in the Introduction.

Response:

Research Questions and Study Objectives have been added to the Introduction. Please check Page 3, Lines 88-95. Three research questions are now explicitly stated to guide the study and frame the interpretation of the results.

3. Title: Remove the final dot at the end of the title.

Response:

The title has been revised accordingly.

Reviewer 3 Report

Comments and Suggestions for Authors

The dataset used is very small, which limits the applicability of the proposal in a generalized environment. The authors are requested to analyze how a larger dataset could be applied.

Since the dataset only includes data from Italy, there is no clear way to assess generalization to other environments with different variables. How do the authors justify this point?

The proposal does not include external validation, which creates uncertainty regarding the reliability of the results. In addition, there may be issues related to overfitting. The authors are requested to analyze this issue and provide a possible solution.

The proposal has limitations regarding disease severity, disease progression, and the inability to reduce the presented problem in clinical environments. How do the authors propose to address these limitations?

The authors are requested to present a trial or evaluation exercise in a real-world environment.

The proposal appears difficult to implement in real-time clinical systems. How do the authors analyze, plan, and design their proposal to make real-world implementation feasible?

Author Response

Reviewer 3

We thank Reviewer for careful evaluation, and critical feedback. We have worked to address all concerns.

Comments and Suggestions for Authors

The dataset used is very small, which limits the applicability of the proposal in a generalized environment. The authors are requested to analyze how a larger dataset could be applied.

Response

We thank the reviewer for this important observation. We acknowledge that the IPVS dataset is relatively small compared with datasets typically used for large-scale machine-learning applications. Consequently, the reported results should be interpreted as a benchmark evaluation within the constraints of the available dataset.

A discussion has been added to the Limitations section explaining that larger datasets would improve model robustness by capturing greater variability in speech characteristics, recording conditions, demographic factors, and disease manifestations. Larger datasets would also enable more reliable estimation of model performance and facilitate the development of more generalized classification models. Future work will focus on combining multiple publicly available PD speech datasets and conducting large-scale evaluations to improve model generalization. Please see page 16-17 lines 531-575.

Since the dataset only includes data from Italy, there is no clear way to assess generalization to other environments with different variables. How do the authors justify this point?

Response

We agree with the reviewer that the use of a single Italian-language dataset limits the ability to assess generalization across different languages, populations, and recording environments. The objective of the present study was to provide a systematic benchmark evaluation using the publicly available IPVS dataset rather than to establish cross-lingual generalizability.

To address this concern, we have expanded the Discussion and Future Work sections to explicitly acknowledge this limitation. We now emphasize that future research should include external validation using datasets collected in different languages and recording conditions. We plan to investigate cross-dataset evaluation between the Italian IPVS corpus and the English-language MDVR-KCL dataset to assess model robustness across linguistic and environmental variations. Please see page 17, lines 540-551 and lines 566-575.

The proposal does not include external validation, which creates uncertainty regarding the reliability of the results. In addition, there may be issues related to overfitting. The authors are requested to analyze this issue and provide a possible solution.

Response

We appreciate the reviewer’s concern regarding external validation and overfitting. We acknowledge that no external validation dataset was available within the scope of the present study. To reduce the risk of overfitting, we employed subject-wise train-test partitioning, ensured strict separation of participants between training and testing subsets, and performed hyperparameter optimization exclusively on the training set using GridSearchCV. Please see page 5-6 lines 205-219

Nevertheless, we agree that external validation represents an important next step for establishing model reliability. We have added a discussion clarifying that future work will include independent validation on external datasets and cross-dataset experiments to evaluate model generalization beyond the IPVS corpus. Please see page 17, lines 566-575

The proposal has limitations regarding disease severity, disease progression, and the inability to reduce the presented problem in clinical environments. How do the authors propose to address these limitations?

Response

The IPVS dataset used in this study provides binary labels (PD versus healthy) and does not contain sufficient longitudinal information to support progression modeling.

We have revised the manuscript to explicitly acknowledge this limitation. Future work will investigate datasets containing clinical severity scores and longitudinal recordings, enabling the development of models for disease staging, progression monitoring, and personalized assessment. Such analyses would provide a more clinically relevant evaluation of speech-based biomarkers for Parkinson’s disease. Please see page 17, lines 553-558 and 566-575

The authors are requested to present a trial or evaluation exercise in a real-world environment.

Response

A real-world clinical trial was beyond the scope of the present study, which was designed as a retrospective analysis of a publicly available research dataset.

To clarify this limitation, we have added a statement in the Discussion section indicating that prospective clinical evaluation is required before deployment in routine healthcare settings. Future work will focus on collecting real-world speech recordings from clinical environments and evaluating model performance under practical operating conditions. Please see page 17 lines 559-565.

The proposal appears difficult to implement in real-time clinical systems. How do the authors analyze, plan, and design their proposal to make real-world implementation feasible?

Response

One advantage of the proposed approach is that it relies on computationally efficient handcrafted acoustic features and conventional machine-learning classifiers rather than resource-intensive deep-learning architectures. Consequently, feature extraction and classification can be performed with relatively low computational requirements.

We have expanded the Discussion section to clarify that the proposed framework is intended as a foundation for future clinical decision-support tools rather than a fully deployable clinical system. Please see page 17 line 559-565. Future development will focus on real-time speech acquisition, automated preprocessing, user-friendly interfaces, and prospective clinical validation to support practical implementation in healthcare environments. Please see page 17 lines 571-575.

We hope that these revisions and clarifications address all concerns and enhance the quality, transparency, and impact of our article and hope that the revised manuscript is now accepted for publication.

I am looking forward to hearing from you, and I would like to thank you in advance for your attention and your time.

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for this revied version of the paper. The goal and RQs are clear now and well addressed. The paper is well-positioned in the SOTA. However, authors should recheck teh references format since there are errors such as reference [13] in which the authors list is wrong (only the last two names should be included). The abstract format should be rechecked. I still believe that it should be a single paragraph (not several ones as it is now).

Comments on the Quality of English Language

It should be revised.

Author Response

We sincerely thank the reviewer for the positive evaluation of the revised manuscript and for the additional suggestions that helped improve its presentation.

Comments and Suggestions for Authors

Thank you for this revied version of the paper. The goal and RQs are clear now and well addressed. The paper is well-positioned in the SOTA. However, authors should recheck teh references format since there are errors such as reference [13] in which the authors list is wrong (only the last two names should be included). The abstract format should be rechecked. I still believe that it should be a single paragraph (not several ones as it is now).

 

Response:

We thank the reviewer for the positive assessment of our revisions and for identifying the remaining issues.

The reference list has been carefully reviewed and corrected, including Reference 13, where author information was inadvertently formatted incorrectly. We have verified all references to ensure consistency with the journal's formatting requirements.

We have also reviewed the abstract formatting and confirmed that it follows the journal guidelines. Now we formatted the abstract into a single paragraph. Please page 1, line 9-32.

We appreciate the reviewer’s careful attention to these details, which have helped improve the quality of the final manuscript.

 

Please see the attached file for more details and other reviewers response. 

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors explain and elaborate on each of the observations

Author Response

We thank Reviewer for careful evaluation, and feedback.

Comments and Suggestions for Authors

The authors explain and elaborate on each of the observations

Response:

We appreciate the reviewer’s positive assessment and are grateful for the constructive comments provided during the review process. The feedback has significantly contributed to improving the clarity, rigor, and overall quality of the manuscript. We sincerely thank the reviewer for the careful re-evaluation of our manuscript.

 

Please see also the attached file for more details of reviewers response. 

Author Response File: Author Response.pdf