Bayesian Opportunities for Brain–Computer Interfaces: Enhancement of the Existing Classification Algorithms and Out-of-Domain Detection

Round 1

Reviewer 1 Report

The authors proposed using Bayesian Neural Network to improve the classification of data collected at BCI experiments. The study is interesting, however, some improvements are needed to make the manuscript suitable for publishing. The introduction needs significant modification by adding more details of the relevant studies. In the current version, very brief explanations of relevant papers are provided.

The clarity of the method section can be improved by adding a chart or block diagram representing the proposed method.

It is not clear why 4Hz and 38 HZ were selected as the cutoff frequencies! while the movement information is acquired by mu and beta rhythms (i.e. 8Hz-30Hz). I expect some changes in the results by removing uninformative rhythms.

The results section also needs significant improvement by adding a clearer interpretation of the results.  

The presentation of references is not consistent in the current draft of the paper (for example references 5,8,15,....). This also should be addressed in the revised version.

-

Author Response

Please see our response to Reviewer 1 in the attached PDF file

Author Response File: Author Response.pdf

Reviewer 2 Report

Major remarks:

- Could you explain why data was bandpass filtered between 4 Hz and 38 Hz?

- What decided the split dataset into two classes: in-domain and out-of-domain?

- Is there a code repository available somewhere?

- Could you add more information about selected values of hyperparameters (eg. the number of epochs, learning rate, batch size, etc.)?

- The results are not convincing. Perhaps it is worth conducting analyses on another dataset as well?

- In some cases, reference to other works is insufficient. In my opinion, some elements need a more detailed description.

Minor remarks:

- At the end of most equations should be a comma or full stop.

- I propose the alphabetical order of references.

Author Response

Please see our response to Reviewer 2 in the attached PDF file

Author Response File: Author Response.pdf

Reviewer 3 Report

Dear Authors,

I have had the opportunity to thoroughly read your manuscript concerning using Bayesian methods in BCI classification tasks. I would like to offer some suggestions and questions that may contribute to enhancing your research article.

1. The methods section describes a complex experimental procedure with significant detail regarding the architectures and parameters used. Please include a flow chart of the methodology, specifically the two experiments conducted, to provide a visual understanding of the experimental design.

2. The discussion mentions the inconsistency in the performance of OOD detectors. Why these inconsistencies occurred, and what features or circumstances led to varying performance? Include this for future works

3.  Could you elaborate on the specific mechanisms or theoretical foundations that led to the higher generalization properties of BNNs compared to deterministic counterparts? How do these findings align/contrast with previous research in the domain? Include this in the discussion.

4.  What are the potential reasons for the Bayesian framework falling short in providing accurate uncertainty estimates in the EEG classification domain? Do you think particular limitations in the data hinder this estimation? Include this in the discussion.

5.  Considering the time complexity, how feasible is the application of BNNs in real-time BCI? What strategies could be employed to overcome the computational challenges of online experiments? Include this in the conclusions.

6.  You mentioned the relevance of OOD data detection in various fields like medicine, robotics, and particle physics. How might the specific methodologies applied in this study be employed in these other domains? Include this in the conclusions

Best Regards,

Author Response

Please see our response to Reviewer 3 in the attached PDF file

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

-

-

Author Response

Thank you for reviewing our manuscript!

Reviewer 2 Report

The authors made a considerable effort to correct and improve their paper in such a short time. They responded to all raised problems and I appreciate this. 

Author Response

Thank you for reviewing our manuscript!

Reviewer 3 Report

After the revisions made, I think the article can be accepted for publication in the current state 

Author Response

Thank you for reviewing our manuscript!