Potential Use of BME Development Kit and Machine Learning Methods for Odor Identification: A Case Study

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents an interesting case study on integrating the BME Development Kit with machine-learning classifiers for food identification based on volatile gas signatures. The manuscript has several flaws that should be addressed.

1. Section 3 currently describes raw TPHG variables (temperature, pressure, humidity, gas resistance) but does not specify how these raw signals are converted into discriminative features. The paper should be extended with a dedicated subsection that lists every engineered feature. Including the exact mathematical expressions or code references will allow readers to reproduce the work and understand why certain features were chosen over others.
2. The manuscript only mentions “BME Development Kit + ESP32” and “BME AI-Studio” at a high level. For full reproducibility, the authors should clarify the exact hardware revision, Python environment, and the random-seed settings. When the Neural Network is trained in BME AI-Studio, the paper must disclose the network topology. For the Random Forest model, please report the number of trees, maximum depth, and impurity criterion.
3. 3. Accuracy alone is insufficient for an imbalanced, four-class problem. The results section should be augmented with macro-averaged precision, recall, F1-score, and the per-class ROC-AUC curves. A single ROC graph (or PR curve) comparing all algorithms would quickly reveal which classes are harder to separate and whether the high accuracy is driven by the majority class. A confusion matrix in raw counts rather than percentages would also help readers gauge absolute misclassification numbers.
4. Additional Technical and Editorial Issues
5. a) The abbreviation “TPHG” is introduced once and then used repeatedly without reminding the reader. Spell it out on first use in the abstract and again in the main text.
6. b) The manuscript occasionally mixes American and British spelling (“analyze” “analyse”). Pick one and apply consistently.
7. c) Many of figures are not clear, such as the labels for the horizontal and vertical axes in Figure 4, the lack of label units in Figure 6, and the confusing layout in Figure 7.

 

Author Response

First and foremost, we would like to express our sincere gratitude for the time and effort devoted to reviewing this manuscript and providing thoughtful suggestions for its improvement. Your comments and feedback are greatly valued. Thank you very much for your constructive feedback and guidance, which have contributed significantly to enhancing the quality of this manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The relevant modification suggestions are as follows:
(1) The article is too long and not concise. Many unimportant tasks take up a large amount of space, and it is recommended to shorten the entire text. As suggested in the previous section on BME and machine learning, it is recommended to compress them appropriately. And the introduction of related algorithms, this part is already very common, there is no need to write too much space.
(2) What is the purpose of the research? Why choose four types of food? Is classification research only aimed at distinguishing four types of samples? What is the reason for the low recognition rate?
(3) The focus of the article is not prominent, and it is not meaningful to simply combine BME data with several machine learning algorithms for classification research. Suggest exploring deeper issues.

Author Response

First and foremost, we would like to express our sincere gratitude for the time and effort devoted to reviewing this manuscript and providing thoughtful suggestions for its improvement. Your comments and feedback are greatly valued. Thank you very much for your constructive feedback and guidance, which have contributed significantly to enhancing the quality of this manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The authors combined low-cost sensors with machine learning to verify the feasibility of electronic noses in food classification, which has practical value for engineering applications in food quality and food quality monitoring. However, the authors need to make improvements or answers to the following questions.

(1) The core method is not original, and similar studies have been reflected in many literatures. If the application and configuration optimization of specific hardware are the only innovation points, then the manuscript is relatively less innovative.

(2) The authors verified the reliability of the conclusions through the combined application of multiple algorithms, data changes and segmentation methods, but there seem to be some issues that the authors need to consider. For example, the authors mentioned the interference of VOCs on sensors, but did not quantify it, which to some extent affected the objectivity of the results. In addition, the samples should consider sample data of fresh food in a corrupted or inedible state.

(3) The authors should add feature importance analysis verification of atmospheric pressure to further enhance the persuasiveness of the conclusions.

(4) The authors explained the accuracy of various methods, but for confusion matrices or certain issues, such as the high misclassification rate of some classes, what is the reason? The authors did not further analyze in depth.

(5) The comparison between different methods also lacks a more in-depth analysis, such as feature space visualization.
(6) The attribution of "Random Forest's improved performance" is too general, and its advantages are not analyzed in combination with the characteristics of electronic nose data (such as high dimensionality and noise).
(7) The introduction to machine learning takes up too much space. The authors can further streamline it and quote the summary statements in the review paper.

Author Response

First and foremost, we would like to express our sincere gratitude for the time and effort devoted to reviewing this manuscript and providing thoughtful suggestions for its improvement. Your comments and feedback are greatly valued. Thank you very much for your constructive feedback and guidance, which have contributed significantly to enhancing the quality of this manuscript.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The authors present a study on odor identification using the BME Development Kit combined with machine learning methods, focusing on classifying four food types. The work contributes to the application of low-cost sensor platforms and machine learning in food quality assessment. However, there are several areas where clarification and improvement are needed.

 

1. There is an inconsistency in the accuracy value of the Neural Network model from BME AI-Studio. Figure 7 shows an accuracy of 90.50%, while the text mentions 90.05%. The authors are requested to verify and correct the data to ensure consistency.

 

2. In Figure 7, the class labels include Portuguese ("COUVE") instead of English. To maintain language consistency throughout the manuscript, it is recommended to replace all Portuguese labels with their corresponding English terms.

 

3. The manuscript highlights the use of the BME Development Kit as a low-cost solution, but lacks a direct comparison with existing electronic nose systems that use dedicated sensor arrays. The authors should supplement such comparisons to clearly demonstrate the advantages of the BME-based approach.

 

Author Response

First and foremost, we would like to express our sincere gratitude for the time and effort devoted to reviewing this manuscript and providing thoughtful suggestions for its improvement. Your comments and feedback are greatly valued. Thank you very much for your constructive feedback and guidance, which have contributed significantly to enhancing the quality of this manuscript.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The author has replied to all my questions.

Author Response

First and foremost, we would like to express our sincere gratitude for the time and effort devoted to reviewing this manuscript and providing thoughtful suggestions for its improvement. Your comments and feedback are greatly valued. Thank you very much for your constructive feedback and guidance, which have greatly contributed to enhancing the quality of this manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

After revision, the article has made significant improvements and there are no additional issues, so it can be considered for publication.

Author Response

First and foremost, we would like to express our sincere gratitude for the time and effort devoted to reviewing this manuscript and providing thoughtful suggestions for its improvement. Your comments and feedback are of great value. Thank you very much for your constructive feedback and guidance, which have greatly contributed to enhancing the quality of this manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

(1) The experiment was conducted under "no environmental control" conditions (e.g., temperature, airflow, humidity, etc.), which is designed to simulate real-world scenarios but lacks control over potential confounding factors, which may lead to high data noise and poor reproducibility. It is recommended that authors provide at least a record of environmental parameters and clarify their potential impacts in the discussion.
(2) The manuscript describes the sample as "approximately balanced", but there are still differences in the sample size of each category (e.g., the Cabbage sample is the largest). Not using stratified sampling or considering category weights in cross-validation may affect the impartiality of model evaluation.
(3) The author should supplement the process and cycle of sensor calibration, whether drift correction is carried out, etc. MOS sensors are susceptible to environmental aging and pollution, which is crucial in long-term applications.
(4) Although the neural network (BME AI-Studio) and traditional ML methods are mentioned, the structure, hyperparameters, and training details of the neural network are not explained, making it difficult for readers to judge the fairness of the neural network compared with the traditional method.
(5) The manuscript directly uses all TPHG variables, and does not perform feature importance analysis or dimensionality reduction (e.g., PCA, LDA), resulting in poor model interpretability, and may introduce redundant or noise features (e.g., air pressure), which the author should supplement.

(6) The performance difference between models (e.g., RF vs kNN) cannot be confirmed by statistical testing (e.g., McNemar test or ANOVA), and the authors must add it.
(7) Although the manuscript mentions several relevant literatures, it is not systematically compared under the same indicators and the same data division method, which is not convincing.

Author Response

First and foremost, we would like to express our sincere gratitude for the time and effort devoted to reviewing this manuscript and providing thoughtful suggestions for its improvement. Your comments and feedback are of great value. Thank you very much for your constructive feedback and guidance, which have greatly contributed to enhancing the quality of this manuscript.

Author Response File: Author Response.pdf