List of Papers By topics Author List
| Paper Info | Reviews | Meta-review | Author Feedback | Post-Rebuttal Meta-reviews |
Authors
Vamsi Kumar, Likith Reddy, Shivam Kumar Sharma, Kamalaker Dadi, Chiranjeevi Yarra, Raju S. Bapi, Srijithesh Rajendran
Abstract
Modeling effective representations using multiple views that
positively influence each other is challenging, and the existing
methods perform poorly on Electroencephalogram (EEG) signals for sleep-staging tasks. In this paper, we propose a novel multi-view self-supervised method (mulEEG) for unsupervised EEG representation learning. Our method attempts to effectively utilize the complementary information available in multiple views to learn better representations. We introduce diverse loss that further encourages complementary information across multiple views. Our method with no access to labels, beats the supervised training while outperforming multi-view baseline methods on transfer learning experiments carried out on sleep-staging tasks. We posit that our method was able to learn better representations by using complementary multi-views.
Link to paper
DOI: https://link.springer.com/chapter/10.1007/978-3-031-16437-8_38
SharedIt: https://rdcu.be/cVRto
Link to the code repository
https://github.com/likith012/mulEEG
Link to the dataset(s)
SleepEDF: https://www.physionet.org/content/sleep-edfx/1.0.0/
SHHS: https://sleepdata.org/datasets/shhs
Reviews
Review #1
- Please describe the contribution of the paper
To combine the temporal and spectral information for representation learning.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The manuscript is well organised.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
The proposed idea is not very reasonable and has not been fully justified in the manuscript.
- Please rate the clarity and organization of this paper
Very Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
It should be reproducible based on the information provided.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
In this manuscript, the authors proposed to combine EEG time series and their spectral representations to enhance the effectiveness of the feature extraction. It has been widely proven that the fusion of multiple features positively contributes to the classification. However, it seems inefficient to combine the EEG time-series and spectral powers. In general, limited information can be mined from EEG time series. I would like to suggest combining spectral powers with other types of features, such as functional connectivity.
Only one channel was used in this study. Why did you use one channel only? More channels could lead to higher performance. In addition, different channels were selected for different datasets. The same channels should be used. If there are not the same channels, the channels located in the adjacent area should be selected.
As the study’s primary purpose is to have a good representation, it is therefore expected to have detailed comparisons in representations. However, only indirect results were shown in the manuscript.
The number of training samples is 31. How did you set a bitch size of 256?
It states “an initial learning rate of 3e-4”. Why did not you use the format of 10e-4?
Typos: e.g., contrastve on Page 5
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
3
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The fusion of differnt kinds of features has been widely adopted. It is quite mature strategy that has been used in the field.
- Number of papers in your stack
4
- What is the ranking of this paper in your review stack?
3
- Reviewer confidence
Very confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Review #3
- Please describe the contribution of the paper
- The proposed method is significant in EEG Domain for training with mult-view self-supervision approach.
- The experiment setup is good and the evaluation process.
- The training aspect is novel by utilizing two views.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- The proposed method is significant in EEG Domain.
- The experiment setup is good and the evaluation process.
- The training aspect is novel by utilizing two views.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
- Did the authors try other techniques for learning except for RESNET?
- Which data augmention bring better result? I don’t see any discussion on this. Please clarity more on this.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
It seems reproducible. Data is available.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
Overall the idea suggested for multi-view seems convincing.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Based on experimental results, I suggest weak acceptance.
- Number of papers in your stack
5
- What is the ranking of this paper in your review stack?
3
- Reviewer confidence
Somewhat Confident
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
5
- [Post rebuttal] Please justify your decision
Not Answered
Review #4
- Please describe the contribution of the paper
The authors propose a multi-view self-supervised method (mulEEG) for unsupervised EEG representation learning. The results show that the proposed method has higher performance.
- Please list the main strengths of the paper; you should write about a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Using complementary information to construct positive pairs in contrastive learning for EEG is novel. The paper is well-written.
- Please list the main weaknesses of the paper. Please provide details, for instance, if you think a method is not novel, explain why and provide a reference to prior work.
More ablation studies about data augmentation are recommended.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Note, that authors have filled out a reproducibility checklist upon submission. Please be aware that authors are not required to meet all criteria on the checklist - for instance, providing code and data is a plus, but not a requirement for acceptance
As the experimental descriptions are enough, I believe this paper is reproducible.
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review: https://conferences.miccai.org/2022/en/REVIEWER-GUIDELINES.html
The author claim that they design an EEG augmentation strategy for multi-view SSL. It is better to show the effectiveness of using augmentations from baselines.
- Rate the paper on a scale of 1-8, 8 being the strongest (8-5: accept; 4-1: reject). Spreading the score helps create a distribution for decision-making
5
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Using complementary information to construct positive pairs in contrastive learning for EEG is novel. The paper is well-written.
- Number of papers in your stack
7
- What is the ranking of this paper in your review stack?
2
- Reviewer confidence
Confident but not absolutely certain
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Not Answered
- [Post rebuttal] Please justify your decision
Not Answered
Primary Meta-Review
- Please provide your assessment of this work, taking into account all reviews. Summarize the key strengths and weaknesses of the paper and justify your recommendation. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Summary & Contribution: This work proposes to use a multi-view self-supervised method called mulEEG for Electroencephalogram (EEG) representation learning. The main motivation of this work is to improve self-supervised methods for EEG as they currently perform poorly, and supervised methods are biased towards the annotator. The authors use a multi-view framework that aims to find a representation from different views. Time-series and spectrogram are use as multiple views of the same EEG signal. Each time-series sample is augmented and converted to a spectrogram. They use a contrastive loss that maximises similarity between augmented samples and minimises similarity with other samples. Evaluation is done on two public datasets (78 + 326 subjects) Results show that the proposed method achieves better performance.
The main contribution of this paper is the use of a multi-view self-supervised method for EEG that learns effective representations.
Key strengths:
- Unsupervised representation learning in EEG is novel and clinical relevant for sleep analysis.
- Strong evaluation of the proposed methodology.
- The idea of using a multi-view approach with a contrastive loss is sound.
Key weaknesses:
- Data split into pretext/train/test is not fully clear. Each EEG from a subject is segmented into non-overlapping 30 seconds samples, and each epoch belongs to one group (pretext/train/test) but it is not clear if all epochs from the same patients belong to the same group (page 3).
- The motivation of some decisions seems to be missing (e.g., number of channels, use of the datasets)
- No statistical analysis is presented in the evaluation of the results
Evaluation & Justification: Reviewers agree that unsupervised representation learning for EEG is novel and interesting. The idea of using data augmentation on the signal and generate the spectrograms is also interesting. However, some questions have been raised on the number of channels used and the data split which need clarification.
If a rebuttal is submitted, please clarify questions raised by reviewer 1 on the use of one channel only. Please also clarify how the epochs were divided into the different datasets. The explanation for the “within dataset” is not clear either, please consider extending the description for this case (e.g., how many cases and which dataset was used, ‘evaluation on train’ what is this?).
- What is the ranking of this paper in your stack? Use a number between 1 (best paper in your stack) and n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
8
Author Feedback
We thank all the reviewers for your constructive feedback. Please find our responses below.
Reviewer#1:
[Q1] Why single EEG channel and different channel names used? [A1] The SleepEDF and SHHS datasets had only 2 EEG channels available in each, and each channel was configured differently in each cohort. In this study, we selected Fpz-Cz channel from SleepEDF and C4-A1 from SHHS dataset. We reported results based on one channel to be consistent with prior work (e.g., see reference [10], [9]). We have also tested using both EEG channels and did not observe significant performance improvement. The two selected channels can be used interchangeably for sleep-stage monitoring [van Sweden et al., Alternative Electrode Placement in (Automatic) Sleep Scoring (Fpz-Cz/Pz-Oz versus C4-A1), Sleep J, 1990].
[Q2] Training samples and batch size [A2] We clarify that “31” in the manuscript refers to # of training subjects, and each subject has a continuous EEG recording, comprising an average of 1500 EEG epochs or samples. Since the number of samples is not small and contrastive loss works better with larger batch size, we chose batch size of 256.
Reviewers #2 and #3:
[Q1] Augmentation discussion [A1] We have experimented with several augmentation strategies with varying strengths of augmentation. Apart from jittering, masking, flipping, and scaling as mentioned in the paper we have also tested random cropping, multi-masking, and randomized bandpass filtering. We have experimented with these augmentations in several combinations and with different strengths of augmentation and reported the one with the best results in the paper. For transparency, we will make the code publicly available to reproduce the results across several augmentation strategies.
Meta-Reviewer:
[Q1] Details on data split [A1] We took extra care to make sure that subject-specific epoch data belongs to only one group within pretext/train/test groups. None of the subject epochs was repeated in two or more groups to avoid data leakage. Data split procedure in detail: First, each subject is randomly assigned to one of the pretext/train/test groups. Second, within that group, each subject’s night-long EEG recording is segmented into 30 non-overlapping epochs/samples. Finally, after group allocation, we evaluated our proposed architecture across two scenarios: within-dataset and transfer learning (across-datasets). Majority of the work done on self-supervised learning on physiological signals (EEG, ECG) generates the pretext/train/test groups within the dataset itself (within-dataset). We kept subjects in the pretext group fairly large and both the train/test groups are combined (only for the within-dataset evaluation) and a 5-fold evaluation is done. Such a data split is not ideal for transfer learning experiments because in reality train/test splits and pretext group would be coming from different datasets, respectively. To represent such scenarios we have also evaluated our model where the pretext group comes from the larger SHHS dataset and the train/test group comes from SleepEDF. Here train/test groups are not combined but all the subjects in the dataset are shuffled for every run with 58/20 subjects selected as train/test groups. We performed 20 such runs (this replaces k-fold evaluation) and report the average performance. We will be adding a table showing the number of subjects and total epochs for each data split group for both the scenarios in the supplementary section to remove confusion.
[Q2] Statistical Analysis: [A2] We thank the reviewer for this suggestion. We have now performed statistical analysis and the results will be included in the manuscript. In transfer learning experiments (Table 1), metrics for our method (Acc: 78.54; kappa: 69.14) are statistically significant versus supervised baseline (Acc: 77.88; kappa: 68.38) with a significant p=0.024 (Acc) and for kappa (p=0.029) using independent samples t-test with unequal variance.
Post-rebuttal Meta-Reviews
Meta-review # 1 (Primary)
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
The main contribution of this work is a multi-view self-supervised method for EEG that learns effective representations.
Key strengths:
- Unsupervised representation learning in EEG is novel and clinically relevant for sleep analysis.
- Strong evaluation of the proposed methodology.
- The idea of using a multi-view approach with a contrastive loss is sound.
Key weaknesses: There are no key weaknesses after rebuttal.
Review comments & Scores: The main concerns of this work were the use of one channel only and details on the data split and dataset.
Rebuttal: The use of a single channel has been clarified in the rebuttal and I do not have any concerns regarding the choice of channel. Authors have also provided clear information regarding the data split ensuring that data from a specific subject was not used in different groups.
Evaluation & Justification: This paper presents a novel an interesting methodology using a multi-view approach with a contrastive loss for unsupervised representation learning in EEG. I believe the evaluation of the proposed method is strong and the authors have addressed successfully my concerns in the rebuttal.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
5
Meta-review #2
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
This paper proposes a novel multi-view self-supervised method for EEG. Following my reading of the paper, reviews, and rebuttal, it seems the authors have addressed most of the concerns. Recommend to accept and ask the authors to reflect the rebuttal points in the paper if finally accepted.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
8
Meta-review #3
- Please provide your assessment of the paper taking all information into account, including rebuttal. Highlight the key strengths and weaknesses of the paper, clarify how you reconciled contrasting review comments and scores, indicate if concerns were successfully addressed in the rebuttal, and provide a clear justification of your decision. If you disagree with some of the (meta)reviewer statements, you can indicate so in your meta-review. Please make sure that the authors, program chairs, and the public can understand the reason for your decision.
This manuscript presented a “multiview” EEG classifier, where the mutlview component comprises spectral and temporal data obtained from a single EEG channel.
Overall, I think this paper has an interesting approach to data fusion for classification of EEG signals. Overall, I found the use of a single EEG channel was a not well justified, and the authors did not justify this well in their rebuttal. But otherwise I think they did a decent job of addressing the methodological concerns on the data augmentation, and training/testing splits. Overall, I think the idea is relatively straight forward, with sensible comparisons and it has interest and relevance for MICCAI.
- After you have reviewed the rebuttal, please provide your final rating based on all reviews and the authors’ rebuttal.
Accept
- What is the rank of this paper among all your rebuttal papers? Use a number between 1/n (best paper in your stack) and n/n (worst paper in your stack of n papers). If this paper is among the bottom 30% of your stack, feel free to use NR (not ranked).
9