Sensor-Based Fault Diagnosis and Prognosis of Neurophysiological States: A Transformer Autoencoder Approach to EEG Monitoring

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript presents a transformer autoencoder framework for analyzing EEG signals across therapy phases, compared against a VAE baseline. The topic is relevant and the idea of using latent representations for monitoring neurophysiological evolution is interesting, especially in low-data clinical settings. The paper is technically coherent and the methodology is described clearly, with appropriate use of autoencoders, attention mechanisms, and latent-space analysis .

However, the level of novelty is a little bit moderate. The contribution mainly consists of applying a transformer-based autoencoder to a small aggregated EEG feature set and comparing it to a VAE. Since the input is limited to 10 features rather than raw temporal EEG, the advantage of transformers is somewhat reduced, and the work appears more as an incremental extension of existing autoencoder-based EEG analysis rather than a fundamentally new approach.

The literature review is generally relevant but lacks strong positioning with respect to recent transformer-based approaches in biomedical signal analysis and disease diagnosis. The paper would benefit from citing and discussing works where transformers are used for clinical decision-making or diagnosis, such as transformer-based EEG analysis surveys (Vafaei & Hosseini, 2025) and works using transformers for Parkinson’s disease detection (Bensefia et al. 2025).

The main weakness lies in the experimental validation. The dataset is extremely small (six subjects), and no rigorous subject-level validation protocol is applied. The results rely heavily on UMAP/t-SNE visualizations and qualitative clustering interpretations, which are not sufficient. Quantitative comparisons between models are limited, and the claims of improved separation and stability are not supported by robust statistical validation across multiple runs. Moreover, the use of terms such as “diagnosis” and “prognosis” is not fully justified, as no predictive modeling or clinical classification is performed.

Overall, the paper is promising but not yet redy for publication. The methodology is sound, but the novelty is limited and the experimental validation is not sufficiently rigorous. I recommend major revision, with stronger literature positioning (including transformer-based disease diagnosis works), clearer claims, and more robust and reproducible evaluation.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This study presents a condition monitoring system for diagnosing and predicting neurophysiological states using electroencephalographic (EEG) signals. Using a deep learning architecture, a baseline variational autoencoder is compared with a transformer-based autoencoder for modeling hidden representations of EEG dynamics. The goal itself merits further consideration. However, I recommend the authors pay attention to the following points.

1. The authors chose an arbitrary paper structure. This is possible, but not in the presentation of the primary research data. Perhaps the article entry guidelines for the specific journal should be reviewed and followed. Also, the bibliography and citations are incorrect.
2. Lines 42-46. The present document is organized into the following sections: Section 2 explores various proposals that use tools similar to the proposed research. Section 3 presents how the data are distributed and the analysis they provide. Section 4 explores the tools used for this research. Section 5 describes the series of activities and tests conducted using the combined use of autoencoders and Transformers and their results. Finally, Section 6 summarizes the key knowledge, the findings obtained, the future lines of research, and the conclusions of the study.
3. The structure of the paper mixes results and methods. These should be described separately.
4. The Introduction and related works do not allow for an assessment of the novelty of the study and the main objective. It would be better to provide a comprehensive Introduction with an analysis of the EEG literature, tied to specific functions and the possible impact of dolphin therapy. EEG assessment over a long period of time (years) raises many questions. The main advantage of EEG is its high temporal resolution (seconds, minutes). This is important for urgent diagnostics or online assessment of cognitive functions.
5. Figures 1, 2, 3, 11, and 16 do not contain information relevant to the article.
6. The subjects are not described at all, although there is specific information about them in the text. Since there are only six subjects, it is necessary to provide as much medical and social background as possible, including their medical history and medication intake. Neurological patients typically take medications on a regular basis.
7. In neurophysiology, it is considered mandatory to submit original EEGs with their descriptions.
8. A discussion in the traditional format is not provided. Here, you should compare your own results with similar data from other laboratories.

 

Based on the above comments, I do not recommend publishing the manuscript in this format.

Author Response

"Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This article lacks innovation, has unreliable experiments, and insufficient clinical support. It is not yet suitable for publication.

1. The comparison with VAE in the paper is unfair: VAE has a 2-dimensional latent space, while Transformer has a 16-dimensional one. The direct difference in dimensions leads to the disparity in clustering effects, and thus cannot prove the architectural superiority.
2. EEG is a time-series signal. In this paper, the frequency band mean is directly used as the static feature input, thus losing the dynamic information of time. Does this contradict the original intention of "neural state monitoring"?
3. In the analysis of the research results, there is a heavy reliance on subjective visualizations using UMAP/t-SNE, while there is a lack of quantitative statistical tests.
4. The title of the paper emphasizes "fault diagnosis/prognosis", but the entire text lacks definitions of faults, abnormal detection, and prediction. It only performs stage clustering, which does not match the title.
5. In this paper, no healthy control group or sham intervention control group was set up, which makes it impossible to rule out the placebo effect.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The revised manuscript shows a clear effort to address the previous concerns, particularly by strengthening the discussion on novelty, expanding the literature review, and clarifying the scope of the claims. 

The authors have appropriately moderated their claims by clarifying that the framework provides qualitative indicators rather than predictive clinical outputs, and by explicitly discussing the limitations related to the small dataset and the lack of subject-level validation. This is a positive and necessary improvement that aligns the conclusions with the actual experimental evidence. However, while these clarifications improve the rigor of the narrative, they do not fundamentally address the underlying limitation of the study, namely the absence of robust quantitative validation. The reliance on visualization-based analysis and the limited statistical evaluation remain significant weaknesses, even if now explicitly acknowledged.

The methodological contribution remains coherent, and the comparison between VAE and transformer autoencoder is reasonably motivated. Nevertheless, the novelty is still incremental, as the approach largely combines established techniques applied to aggregated EEG features rather than raw temporal signals, which limits the advantage of the transformer architecture. The added justification clarifies the intent but does not substantially elevate the originality of the contribution.

In summary, the revision has improved the clarity, positioning, and honesty of the manuscript, and the authors have responded constructively to the reviewer’s feedback. However, the experimental validation is still insufficient to fully support the claims, and the contribution remains moderate in novelty. The paper is improved but would still benefit from stronger quantitative validation and more rigorous experimental design before being considered fully ready for publication.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The paper has been significantly revised in structure and content. Since the manuscript discusses a rather serious neurological pathology, this line of discussion should be strengthened in the Discussion section, with a comparison to similar studies. A separate Limitations section should be created. The Conclusion can be shortened, retaining only the significant findings.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

I have no comments.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf