Demystifying Mental Health by Decoding Facial Action Unit Sequences
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe paper presents a methodology for estimating emotional states from face images of healthy subjects. The proposed approach consists of a preprocessing pipeline, an extraction step of facial action units, and a convolutional neural network for the classification of facial microexpressions labeled with emotions.
The presented approach is interesting and the validation conducted on literature datasets seems promising. I suggest that the authors respond to a number of comments to improve the presentation of the work:
- The abstract should contain some summary results of the proposed approach. Already from the abstract, the reader should understand whether the new approach proposed in this paper is comparable to the state of the art.
- Lines 23-60 present the problem in an overly discursive and extensive manner. This reviewer suggests summarising and going straight to the problem to be addressed.
- In which application scenarios can this technology play a role? What are the experimental, clinical, and assistive scenarios in which the authors envision their integrated technology? Can motor rehabilitation, as demonstrated in 10.1109/JSYST.2023.3317504, exploit such approaches to make robots aware of a person's internal state?
- In the state of the art, I suggest the authors cite recent work on facial expression recognition such as: 10.3390/s23208376, 10.1007/s00521-021-06012-8.
- the approach described in lines 153-158 does not have the citation.
- enlarge the font of the wording in Figure 2 to make it more readable.
- section 3.2 appears to be too fragmented. I suggest the authors reduce the number of indentations and rephrase the text to make it easier to read.
- Clustering analysis is not mentioned earlier in the text. The authors should clarify its function and usefulness at the outset.
- The input dimension of the model is not clear. 48x48 is the size in pixels. How is it possible that 512 channels are used for each pixel? This reviewer needs more detail in order to understand. Which preprocessing expanded the information from 3 channels (RGB) to 512?
- Is the architecture presented in lines 338-347 new, presented by the authors, or does it take inspiration from literature? The authors should clarify this.
- Lines 362-386 present the metrics used to validate the approach. This part deserves a section within Methods, before the Results section. Similarly, lines 396-401, are explanations of the method used, not results.
Author Response
Thanks for the valuable comments; please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper introduces a solution to mental health diagnosis by classifying facial actions in videos. Sequential modeling is integrated into classification framework and the public datasets SAMM and CASME are used to examine the effectiveness of the proposed solution. CNN is employed to perform deep learning with SAMM and CASME for classification. The method achieves 95.62% accuracy, 93.56% precision, 95.62% recall, and 91.78 F1-score on testing set.
This paper critiques the lack of novelty and limited contribution to data mining exhibited within its contents. While it attempts to propose an approach aimed at addressing mental health analysis issues, it falls short in several aspects. Notably, there is an absence of comparison with existing methodologies. This is critical to align this work with existing approaches, which determines if this work is unique and provides valuable insights into the effectiveness. Furthermore, the analysis of the dataset appears to lack depth, failing to reveal useful questions or insights for further investigation. The discussion section, though present, is hurried and offers scant information regarding the practical implications of the classification results in medical diagnosis and practice. Additionally, the absence of a processing framework figure poses a hindrance to comprehension, as it obstructs understanding of the workflow described in the paper. In summary, while the paper develops a solution, it needs substantive revisions and enhancements to fully realize its potential contribution to the field of data mining and mental health analysis.
Comments on the Quality of English LanguageThe general quality of the content is well-written and articulate, but minor editing is still needed.
Author Response
Thanks for the valuable comments; please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsNo other comments.
Author Response
Thank you for your time and support in reviewing this manuscript. We appreciate it !!
Reviewer 2 Report
Comments and Suggestions for AuthorsThank you for your detailed response, which has been helpful in addressing some of the questions. However, I still have the following concerns regarding the current manuscript.
1. The algorithm is the only aspect of this research showing limited contribution. Both the data and the model are existing and conventional. To enhance this, I would expect a more informative discussion on the results, including statistical analysis such as an ablation study among different methods. While the practical applications are sufficient, they need slight reduction for better context balance.
2. Technical details need to be clarified. CNN is not a standard architecture. The experiment needs the information about layer numbers, kernel size, activation functions, and sequential stacking order. This information applies to all neural networks and algorithms.
3. All figures need clearer presentation with larger fonts, higher resolution, and a well-designed scale. The framework (Figure 2) requires significant improvement in presentation to achieve a more elegant depiction, as it is the most important figure in this paper.
Comments on the Quality of English LanguageMinor language editing is required for improvement.
Author Response
Please see the attacehed.
Author Response File: Author Response.pdf
Round 3
Reviewer 2 Report
Comments and Suggestions for Authors1. Please review the details of the CNN architecture. The layer summary appears to contain incorrect numbers, specifically the change from 64 to 32 after the MaxPooling2D layer.
2. The Decision Tree model fails, indicating either a potential issue with the implementation or the need for an explanation behind this observation. Furthermore, a more comprehensive discussion on the ablation study is necessary to understand the variability of this method across different parameter settings.
3. The discussion section needs to be more coherently organized. Irrelevant content that is not related to medical mental health should be reduced, and more relevant information focusing on mental health analysis applications should be discussed in greater detail.
Comments on the Quality of English LanguageMinor editing is still needed.
Author Response
Please see the attachment
Author Response File: Author Response.pdf
Round 4
Reviewer 2 Report
Comments and Suggestions for AuthorsIt addressed my concerns. I think this paper could be further revised for publication.
Comments on the Quality of English LanguageMinor editing is needed for manuscript improvement.
Author Response
Reviewer #:
- It addressed my concerns. I think this paper could be further revised for publication.
Response: We are thankful to the reviewer for providing us this valuable guidance. As per your suggestion the manuscript has undergone revision. We have thoroughly checked the entire manuscript and made required modifications to the grammar and spelling.
