List of Papers Browse by Subject Areas Author List
Abstract
Electroencephalography (EEG) based automated seizure classification can significantly ameliorate seizure diagnosis and treatment. However, the intra- and inter- subject variability in EEG data make it a challenging task. Especially, a model trained on data from multiple subjects typically degenerates when applied to new subjects. In this study, we propose an attention based deep convolutional neural network with domain adaption to tackle these issues. The model is able to learn domain-invariant temporal-spatial-spectral (TSS) features by jointly optimizing a feature extractor, a seizure classifier and a domain discriminator. The feature extractor extracts multi-level TSS features by an attention module. The domain discriminator is designed to determine which domain, i.e., source or target, the features come from. With a gradient reversal layer, it allows extraction of domain-invariant features. Thus, the classifier is able to give accurate prediction for unseen subjects by leveraging knowledge learned from the source domain. We evaluated our approach using the Temple University Hospital EEG Seizure Corpus (TUSZ) v1.5.2. Results demonstrate that the proposed approach achieves the state-of-the-art performance on seizure classification. The code is available at https://github.com/Dondlut/EEG_DOMAIN.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2024/paper/1802_paper.pdf
SharedIt Link: pending
SpringerLink (DOI): pending
Supplementary Material: https://papers.miccai.org/miccai-2024/supp/1802_supp.pdf
Link to the Code Repository
https://github.com/Dondlut/EEG_DOMAIN
Link to the Dataset(s)
https://isip.piconepress.com/projects/nedc/html/tuh_eeg/
BibTex
@InProceedings{Fan_ADomain_MICCAI2024,
author = { Fan, Xiaoya and Xu, Pengzhi and Zhao, Qi and Hao, Chenru and Zhao, Zheng and Wang, Zhong},
title = { { A Domain Adaption Approach for EEG-based Automated Seizure Classification with Temporal-Spatial-Spectral Attention } },
booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2024},
year = {2024},
publisher = {Springer Nature Switzerland},
volume = {LNCS 15005},
month = {October},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The main contribution of this paper lies in proposing a novel method called DA-ATSS for automated EEG-based seizure classification. This method combines domain adaptation techniques with temporal-spatial-spectral (TSS) attention mechanism to address the cross-subject and cross-domain EEG classification challenges. This research provides an effective solution applicable to automated seizure classification in clinical EEG data and demonstrates significant performance improvements over other baseline methods in experiments.
- 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.
This paper employs an adversarial domain adaptation approach to mitigate inter-individual differences, thereby slightly improving the accuracy of seizure detection.
- 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 paper does not provide a clear explanation of the architecture of the domain adaptation discriminator.
- Why is the architecture of the domain adaptation discriminator the same as that of the seizure classification network? Does adversarial learning require fixing the weights of one side during training?
- The loss function values for seizure classification and domain discriminator during training should be displayed/plotted.
- The evaluation metrics during model training should also be included in Table 2 because the authors need to emphasize the issue of overfitting, which necessitates the use of domain adaptation to overcome it. In simple terms, the experimental results tabulated do not effectively highlight the true benefits of domain adaptation. While Table 2 shows that the channel or spatial attention mechanism effectively improves overall accuracy, this aspect is not an original method proposed by the authors. Therefore, the substantive academic contribution of this paper is not particularly high.
- Please rate the clarity and organization of this paper
Poor
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
No
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
- The paper does not provide a clear explanation of the architecture of the domain adaptation discriminator.
- Why is the architecture of the domain adaptation discriminator the same as that of the seizure classification network? Does adversarial learning require fixing the weights of one side during training?
- The loss function values for seizure classification and domain discriminator during training should be displayed/plotted.
- The evaluation metrics during model training should also be included in Table 2 because the authors need to emphasize the issue of overfitting, which necessitates the use of domain adaptation to overcome it. In simple terms, the experimental results tabulated do not effectively highlight the true benefits of domain adaptation. While Table 2 shows that the channel or spatial attention mechanism effectively improves overall accuracy, this aspect is not an original method proposed by the authors. Therefore, the substantive academic contribution of this paper is not particularly high.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making
Reject — should be rejected, independent of rebuttal (2)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
In the issues of domain adaptation, the authors need to emphasize the issue of overfitting, which necessitates the use of domain adaptation to overcome it. In simple terms, the experimental results tabulated do not effectively highlight the true benefits of domain adaptation. While Table 2 shows that the channel or spatial attention mechanism effectively improves overall accuracy, this aspect is not an original method proposed by the authors. Therefore, the substantive academic contribution of this paper is not particularly high.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Review #2
- Please describe the contribution of the paper
The paper makes the following contributions to the field of EEG-based seizure classification: 1.The paper proposes a novel approach that employs domain adaptation to address the challenge of inter-subject variability in EEG data for seizure classification.
- The model incorporates an attention mechanism that focuses on the most discriminative timestamps, sensor locations, and frequency bands within the EEG data.
- 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 paper introduces a domain adaptation approach that is specifically tailored for EEG-based seizure classification. The use of a gradient reversal layer to learn domain-invariant features is a novel contribution that addresses the challenge of subject variability in EEG data. This approach is particularly interesting because it allows the model to generalize well across different subjects, which is crucial for clinical applicability.
-
The incorporation of the TSS attention mechanism is another strength of the paper. By focusing on the most informative aspects of the EEG signals, the model can more effectively distinguish between different seizure types. The attention mechanism is novel in the context of seizure classification, and it provides a way to automatically identify and leverage the most relevant features for classification.
-
The paper includes a thorough ablation study that provides insights into the contribution of different components of the model. This is a strength because it not only validates the design choices but also helps readers understand the importance of each part of the model, such as the attention blocks and the domain adaptation approach.
-
The paper’s methodological approach is sound, with a clear problem statement, a well-defined model, and a rigorous experimental setup. This methodological rigor is a strength that lends credibility to the results and conclusions of the study.
-
- 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 model’s performance is evaluated on a single dataset, TUSZ v1.5.2. While this is a standard benchmark, it may not fully represent the variability seen across different datasets or real-world scenarios. The model’s performance on other datasets could vary, which might be a limitation.
-
While the attention mechanism is a strength, the paper could benefit from a more detailed analysis of what the attention blocks are focusing on and how this relates to known seizure characteristics. Increased interpretability could aid clinical acceptance.
-
The paper could benefit from a more comprehensive discussion of its limitations, including potential sources of bias, the robustness of the model to noise and artifacts in EEG signals, and any assumptions made in the model’s design.
-
The experimental results reported in the paper are significantly related to the division of the dataset. A cross-validation approach could be employed, or alternatively, the same dataset partition utilized by the current state-of-the-art (SOTA) methods could be applied for performance comparison.
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
No
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
The paper presents a novel approach for automated seizure classification using EEG data. The authors propose a deep convolutional neural network with domain adaptation and a temporal-spatial-spectral (TSS) attention mechanism. The model aims to address the challenge of subject variability in EEG data by learning domain-invariant TSS features. The TSS attention module allows the model to focus on the most significant EEG features for seizure classification. The authors evaluate their model on the TUSZ dataset and demonstrate state-of-the-art performance in seizure classification tasks. The paper also includes an ablation study to validate the effectiveness of the attention blocks and domain adaptation approach.
Major Strengths 1.Innovative Approach: The integration of domain adaptation with TSS attention is a novel contribution that significantly improves the model’s ability to generalize across subjects.
Clarity and Presentation Organization: The paper is well-structured, with a clear introduction, related work, methodology, experiments, and conclusion. Clarity: The authors present their ideas and results in a clear and understandable manner. The use of figures and tables enhances the readability of the paper.
Other Comments:
- Provide a more detailed analysis of what the attention mechanism is focusing on and how this relates to known seizure characteristics.
2.The methods compared in Table 2 are re-implemented existing methods by the authors, which may not provide a direct comparison. It is necessary to include a comparison with the most recent SOTA methods in the task of seizure classification for a more intuitive assessment. Furthermore, it would be beneficial to evaluate the model’s performance on additional datasets to further demonstrate its generalizability across different patient populations and recording conditions.
- Include a more comprehensive discussion of the limitations, including potential sources of bias and the robustness of the model.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My recommendation is based on the paper’s innovative approach, and empirical results. The use of domain adaptation and TSS attention mechanism to address subject variability in EEG data is a significant contribution. However, the lack of comparison with the latest SOTA methods are area for improvement.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
Accept — should be accepted, independent of rebuttal (5)
- [Post rebuttal] Please justify your decision
Based on the author’s response, the revised manuscript, taking into account the suggestions for revision, will be an excellent piece of work.
Review #3
- Please describe the contribution of the paper
The contribution of this paper lies in proposing an attention-based deep convolutional neural network with domain adaptation for EEG-based automated seizure classification. This model addresses the challenge of intra- and inter-subject variability in EEG data, which often leads to performance degradation when applied to new subjects. By jointly optimizing a feature extractor, a seizure classifier, and a domain discriminator, the model learns domain-invariant temporal-spatial-spectral (TSS) features.
- 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 problem is important.
- The proposed method is well-explained.
- The experiments are convincing.
- Figures are appropriate.
- 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 related work section must be enhanced. • The problem definition is not formal enough. Authors should give a clear formal definition of the problem. • Some improvements are needed in the description of the method. • A novel solution is presented but it is important to better explain the design decisions (e.g. why the solution is designed like that) • Some text must be added to discuss future work or research opportunities.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not provide sufficient information for reproducibility.
- Do you have any additional comments regarding the paper’s reproducibility?
NA
- Please provide detailed and constructive comments for the authors. Please also refer to our Reviewer’s guide on what makes a good review. Pay specific attention to the different assessment criteria for the different paper categories (MIC, CAI, Clinical Translation of Methodology, Health Equity): https://conferences.miccai.org/2024/en/REVIEWER-GUIDELINES.html
• The related work section must be enhanced. • The problem definition is not formal enough. Authors should give a clear formal definition of the problem. • Some improvements are needed in the description of the method. • A novel solution is presented but it is important to better explain the design decisions (e.g. why the solution is designed like that) • Some text must be added to discuss future work or research opportunities.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making
Weak Accept — could be accepted, dependent on rebuttal (4)
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The related work section needs enhancement to provide a comprehensive review. The problem definition lacks clarity and requires a formal definition. Method description needs improvements for reproducibility. Design decisions should be explained to justify the solution’s design. Lastly, future work discussions are needed to highlight research opportunities.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the author’s rebuttal, state your overall opinion of the paper if it has been changed
N/A
- [Post rebuttal] Please justify your decision
N/A
Author Feedback
We appreciate all reviewers’ (R1,3,4) constructive comments. We propose a novel approach based on domain adaptation (DA) to address inter-subject variability issue in EEG for seizure type classification. Moreover, an attention mechanism is proposed to extract temporal-spatial-spectral features from EEG. The reviewers found our work interesting and novel (R1,4), with convincing experiments (R4), crucial for clinical applicability (R1). All reviewers recognized the contribution of our work. We elaborate on the comments below. We will revise our manuscript as requested and release our code if accepted. R1: 10.C2&6.1,4: Model evaluation: Evaluating on more datasets would more comprehensively evaluate the model. But annotations of multiple seizure types are not provided in other public datasets. In fact, it is common to use TUSZ as the only dataset for seizure type classification. We did train all models with the same dataset partition, as suggested (6-4). Among these models, Wavelet+lightGBM (2022), CNN (2021), GNN (2022) are recent SOTA models in seizure type classification. CNN+Attention and Transformer-based are attention-based methods for EEG classification. These models are depicted in section 4.2. 10.C1&6.2: Analysis of attention: We used Grad-CAM to interpretate the model’s decision. We found high overlap between sensors attended by model and annotations. We did not show these results due to page limits. We believe it is reasonable for an 8-page single-column conference paper. We can provide as supplementary materials if permits. 10.C3&6.3: Discussion of limitations: We will discuss the source of bias and model robustness as requested. R3: 10.1,2&6.1,2: Domain discriminator architecture and training: The feature extractor, domain discriminator and seizure classifier are jointly optimized. The domain discriminator and seizure classifier are two simple classification heads that take the output of the feature extractor as input and perform two different classification tasks. It is common to use similar architecture, e.g., fc layer. Here, we use a pooling layer, and a conv layer with size 1 (the number of kernels equals to the number of classes, i.e., 2 for domain discriminator, 4 for seizure classifier), which is equivalent to a fc layer. Please see “Seizure Classifier and Domain Discriminator” of section 3.2. We use a gradient reversal layer (GRL) between the feature extractor and domain discriminator such that the feature extractor is optimized to misclassify the domains (learn domain-invariant features) while being able to classify seizures. This is different from the adversarial training in GANs. With GRL, we do not train the seizure classifier and domain discriminator alternatively. 10.3,4&6.3,4: Overfitting and using existing attention blocks: We are not able to show the loss curve during training due to rebuttal policies. The acc on training set is 93% (lower on test set since they are unseen during training). But we argue that with DA the model performs better on test set, with a much higher weighted precision (88.2 vs. 84.4), showing the benefits of DA. Please also see the confusion matrix in Fig. 3. As the reviewer pointed out, the two attention blocks in our work are inspired by ECAnet and spatial attention. Although slight modifications are made, they are not entirely novel. However, we combine them to extract temporal-spatial-spectral features from EEG, which is novel. R4: 10.4&6.4: Better explain design decisions: First, seizures exhibit highly complex temporal dynamics, and vary in temporal evolution. Therefore, we designed a temporal attention to focus on the discriminative time stamps. Second, different seizure types present highly complex spectral distributions over brain. This motivates us to design a spatial-spectral attention. Third, we use DA to relief the inter-subject variability issue. We will explain this more clearly. 10.1-3,5: Manuscript should be enhanced: we will revise it very carefully.
Meta-Review
Meta-review #1
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
N/A
- 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).
N/A
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
The rebuttal does a good job in addressing the concerns noted by the reviewers.
- 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).
The rebuttal does a good job in addressing the concerns noted by the reviewers.