Review Reports - A Deep Learning-Based Decision Support System for Sensory Evaluation: A Predictive Framework for Functional Product Taste Assessment in Neuromarketing

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study attempts to integrate consumer neuroscience and neuromarketing using multivariate methods (PCA and deep learning neural networks) to analyze consumer neural responses to functional products, which demonstrates good novelty and practical value.

However, I found an confusing point makes me hesitate to endorse its acceptance.

In the introduction, the authors claimed "In a study involving 83 panelists aged 20 to 29 years (median = 25), brain activity and facial expressions were recorded using EEG while tasting samples. The study, which ensured prior consent, consisted of 43.7% women and 56.3% men."

However, my calculation suggested this number seems to be impossible. if there are 36 women, it should be 36/83=0.4337 (around 43.4%), if there are 37 women, it should be 37/83=0.4457 (around 44.6%), likely, if there are 47 man, it should be 47/83=0.5662 (around 56.6%), if there are 46 man, it should be 46/83=0.5542 (around 55.4%).

Thus, in conclusion, for a cohort containing 83 participant, there is no way to containing 43.7% women and 56.3% men.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper investigates the relationship between consumer neuroscience and neuromarketing using multivariate methods, including PCA and deep learning, to analyze EEG responses of 83 participants during functional food decision-making. The results identify key neural markers—particularly low beta and gamma bands, along with attention and meditation levels. While the idea is interesting, the paper itself looks raw and should be improved.

What are the main results of this paper? It should be stated explicitly. What is proposed - a neural network architecture? a data preprocessing method with a use of PCA?
It would be beneficial to see the training information, for example loss vs epoch plots, etc. given a small number of samples
The more detailed comparative study is required as currently it's not well structured. At least there should be an overall table where your architecture compares with three other models.
There many minor errors
1. Figure 5 should be improved as the names are overlaid and can't be read
2. Line 267 and 278 - repetition?
3. Fig 8 and Tables 1 and 2 - what is the difference between them?
4. Line 424 - a sentence is meaningless
5. Line 455 - "they proposed" - who proposed?
6. And so on, the paper should be proofread with regard to the style and the structure. There are too many references to external parts of the work; as a result, the paper does not appear to be self-sustaining.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript proposes a neuromarketing decision-support framework that integrates EEG signals, principal component analysis (PCA), and deep convolutional neural networks (DCNNs) for predicting taste preferences of functional foods. However, the study still shows notable shortcomings in terms of model generalization ability, comparative experimental design, statistical significance validation, and objectivity in result interpretation.

1、The manuscript presents an engineering-level integration of EEG, PCA, and DCNN, but fails to clearly articulate the substantive methodological innovations compared with existing neuromarketing studies. The authors are encouraged to explicitly clarify the novelty of their approach in the Introduction.

2、The study is based on data from only 83 participants, which is relatively limited for training deep learning models and may pose a risk of overfitting. The authors should further discuss this limitation and its potential impact on model generalizability.

3、A large number of images are used for model training; however, it remains unclear whether these images and the corresponding EEG samples were strictly separated at the subject level. The authors should clarify whether subject-wise independence was ensured during data partitioning.

4、After incorporating the β and γ frequency bands, the model accuracy increases dramatically from approximately 0.72 to 0.97. This improvement appears unusually substantial. The authors are advised to include statistical significance tests or repeated experiments to rule out potential information leakage.

5、The current experiments only compare scenarios with and without the inclusion of β/γ bands, without benchmarking against alternative models. The authors are encouraged to include baseline model comparisons to better contextualize the reported performance gains.

6、Figure 5 contains overlapping text elements that affect readability. The authors should revise the figure formatting accordingly.

7、The evaluation metrics mainly rely on Accuracy, F1-score, and ROC. The authors are encouraged to additionally report metrics such as AUC and Balanced Accuracy to enhance the robustness of the conclusions.

8、The neurocognitive interpretations of the δ, θ, β, and γ frequency bands are largely based on inferential descriptions. The authors are encouraged to support these interpretations with additional relevant literature on neural mechanisms.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The research sample comprises 83 participants aged 20-29. Why was this age range chosen? Is there a scientific rationale behind it? Does the sample include any special groups, such as pregnant women or individuals with chronic diseases.
Although it is mentioned that the dataset can be downloaded via Google Drive, no detailed documentation on the dataset is provided, including key information such as data format, variable definitions, preprocessing steps, and methods for handling missing values, which may affect the replication and verification of results by other researchers. Furthermore, the statistical basis for determining the sample size is not stated, and there is a lack of argumentation on whether the sample size of 83 participants meets the requirements for model training and stability of results. Details of data collection need to be supplemented.
The specific process for participants to taste functional foods is not clearly stated, such as the sample size for each tasting, whether the specific duration of tasting intervals is set based on scientific evidence, and whether irrelevant variables such as the participants' dietary status before the experiment (e.g., fasting or full) are controlled. These factors may affect taste perception and electroencephalogram (EEG) signals, thereby interfering with the research results. Details of the experimental procedure need to be supplemented.
In the study, it was mentioned that eight functional foods were tested, but key information such as specific types, health claims, and ingredient composition of these foods were not clearly defined. Differences between different functional foods may affect consumers' EEG responses and preferences, and the study did not control for this variable, which may lead to confusion in the results. Need to add some details about functional foods.
The text size in the chart is not uniform, it is recommended to adjust it, such as Figs. 7 and 9.
Is "F1" in line 338 and "f1" in line 373 the same variable?
The reference format is not consistent, it is recommended to make adjustments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 5 Report

Comments and Suggestions for Authors

The manuscript addresses an interesting and timely topic by combining EEG-based neuromarketing with deep learning for sensory evaluation of functional foods. The study is generally well structured and the experimental framework is clearly motivated. However, several minor issues should be addressed to improve clarity, methodological transparency, and scientific rigor.

The manuscript reports 83 participants, 784 EEG tests, 39,143 product images, and 119,216 facial expressions. While these numbers are informative, the relationship between participants, EEG recordings, images, and labels is not sufficiently clear. A short table or schematic explicitly showing how many samples per participant were used for model training and testing would improve reproducibility.
The term functional product is used throughout the manuscript, but the exact nature and composition of the tested food samples are not described in sufficient detail. Please clarify whether these were commercial products or experimental formulations, and briefly list their main functional characteristics (e.g., added nutrients, bioactive compounds).
In Section 2.3.3, PCA results are discussed in depth, including correlation values between EEG bands. However, it is not clearly stated:
- how many principal components were retained,
- what cumulative variance they explain, and
- whether PCA was applied per participant or on pooled data
Since EEG signals, facial images, and product images originate from the same participants, it should be clarified whether data splitting (train/validation/test) was performed at the participant level or at the sample level. Participant-level splitting is strongly recommended to avoid optimistic bias and data leakage.
The performance reported when including β and γ bands (F1-score ≈ 0.97, AUC ≈ 0.95) is remarkably high. A brief discussion addressing potential overfitting, class imbalance effects, or task simplicity would be appropriate, especially given the comparatively modest performance of the model without EEG band weighting.
The ethics section is detailed, but it would benefit from explicitly stating the name of the approving institutional body and whether an approval code or reference number exists, in line with standard journal requirements.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed my comments and I will forward it to the editor for the decision.

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for answering my remarks. I believe they have helped to improve the paper.