EvoBrain: Dynamic Multi-Channel EEG Graph Modeling for Time-Evolving Brain Networks

We appreciate Reviewer w1Ha for recognizing the paper organization, clear contributions, experiments and solid theoretical grounding. We appreciate your time and constructive comments. We hope our clarifications below can address your concerns and raise your rating of the paper.

W1: Two different claims.
As you correctly pointed out, our contributions include two distinct components: (1) a dynamic graph creation and (2) a time-then-graph model. All three model variants (graph-then-time, time-and-graph, and time-then-graph) are compatible with either static or dynamic graph inputs.
To ensure a fair comparison, we applied our dynamic graph structure to all GNN baselines across three model variants, as shown in Figure 3. Our time-then-graph model consistently outperforms the others even when all models use the static or our dynamic graph structures. Moreover, applying our dynamic graph to baseline architectures also improves their performance.

W2: No traditional ML model was used.
We have conducted experiments using Random Forests and SVM on the 12-second seizure prediction task. The AUROC scores were 0.778 for Random Forests and 0.765 for SVM. We will include all results of these traditional baselines in the final version.

W3: Other GNN models.
Thank you for your comments. We conducted additional experiments replacing the GCN layers with a more expressive Graph Isomorphism Network (GIN). The GCN layers contain 24,441 parameters, whereas the GIN layers increase the parameter count to 39,016. In the 12-second seizure detection task, GIN achieved an AUROC of 87.2, which is slightly lower than the 87.7 obtained by the GCN. The excessive model complexity may overfit to transient or noisy patterns rather than capturing stable dynamics.
Based on our experiments, we observed that how different components are integrated (i.e., time-then-graph) is more critical for effective dynamic EEG modeling rather than the complexity of individual modules. We will include these additional findings in the final version of the paper.

S1: Typo.
Thank you for your suggestion.

S2: Nascent.
Thank you for your thoughtful comment regarding the use of the term "nascent". Our intention was to emphasize that while there have been emerging efforts to model temporal signals and graph structures jointly, the modeling of their interactions remains at an early and underdeveloped stage. These methods are often exploratory and lack consistent theoretical or empirical grounding, leading to unstable performance across tasks.

S3: Threshold to rank the correlations.
In our experiment, the value of τ (number of top neighbors) is set to 3 following prior work on EEG functional connectivity modeling. We additionally conducted experiments with τ = 5 and τ = 7, and observed less than 1% variation in performance, suggesting that most connections beyond the top-3 are not informative. We will include these results and relevant discussion in the final version.

S4: Sparse graphs.
Thank you for the discussion. By retaining only the top-τ most correlated neighbors for each node, we eliminate weak or noisy correlations that may not reflect meaningful functional connectivity. This sparsification reduces redundant or task-irrelevant information and improves both interpretability and performance.

S5: Variation in running times.
Thank you for your comments, and we will report the variation across multiple runs in the final version.

Q1: Ablation of GCN.
While the two-stream Mamba effectively captures the temporal dynamics of node activity and edge strength independently, it lacks mechanisms to jointly model the interactions between node representations and their evolving connectivity patterns. Given that seizures are a network-level phenomenon, it is essential to capture not only which nodes are active, but also how their activity propagates through dynamic, task-relevant connections. The GCN layer explicitly fuses node features with their neighboring context, enabling the model to learn spatiotemporal patterns such as synchronized activation or pathological propagation across brain regions.
From our theoretical perspective of time-then-graph modeling, GCN complements the two-stream backbone by capturing cross-modal interactions between node activity and edge-defined structure.
We conducted an ablation study, removing the GCN layer and concatenating node and edge features, then applying a linear classifier. This setup yielded an AUROC of 84.7 on the 12-second seizure detection task, which is notably lower than the 87.7 achieved with the GCN layer. We will include this ablation along with the GIN comparison in the final version.

Q2: Seizure prediction.
Thank you for the constructive comments. Compared to seizure detection, the prediction task holds greater clinical value, as it enables early intervention by clinicians. Seizure prediction is inherently more difficult than detection, as it requires identifying subtle pre-ictal patterns rather than overt seizure activity. Longer prediction windows (such as 60s vs. 20s) contain less seizure-indicative information, making the task even more challenging and demanding greater model capacity.
To the best of our knowledge, this is the first work to conduct seizure prediction on the large-scale TUSZ dataset, which involves a substantial number of patients and supports long prediction windows up to 60 seconds before seizure onset. This setting is more challenging than those considered in most prior studies, which typically focus on shorter prediction windows. We will address these points in the experimental setup section of the revision.

Q3: Imbalanced dataset.
Thank you for your suggestion. To ensure a fair evaluation, we did not apply techniques such as weighted loss functions. Instead, we focused on balanced and robust metrics, F1 score, and AUROC. We will include sensitivity and specificity in the appendix in the final version.

Limitations and negative social impacts.
Thank you for pointing out this important aspects. We will explicitly state these concerns as part of the limitations and broader societal impacts in the final version. Below is our intended discussion:

In seizure prediction, where pre-ictal patterns are typically weaker and more spatially diffuse. While EvoBrain achieves the best performance among lightweight GNN-based models, LaBraM benefits from having approximately 30 times more parameters and pretraining on large-scale multiple EEG corpora, which enhances its ability to generalize under limited and noisy pre-seizure data conditions. While we focus our evaluation on the seizure task, which is particularly critical and life-threatening among various EEG applications, generalization to other tasks remains a limitation of our work.
Regarding potential negative societal impacts, we recognize key risks such as bias and system malfunction. Specifically, models trained on EEG data from specific demographic groups may exhibit biased performance when applied to broader populations, potentially leading to unequal diagnostic accuracy. Additionally, miscalibrated early seizure prediction could lead to false alarms, causing unnecessary interventions or patient distress. These challenges highlight the importance of incorporating fairness assessments, demographic audits, and human-in-the-loop strategies in future development and deployment stages.

Dear Reviewer w1Ha,

We are very happy to hear that we were able to address most of your concerns, and we sincerely thank you again for your recognition. Below, we provide further responses to your remaining two questions:

(1) Please kindly focus on the purple bars in Figure 3. All models, including our EvoBrain, are built on top of a static graph structure. Our method achieves the best performance, with the highest bar in both 12s detection (subfigure a) and 60s detection (subfigure b). This clearly demonstrates that our time-then-graph outperforms the others.

(2) We apologize for the confusion. Yes, we are the first to conduct 60-second early prediction on TUSZ. We truly appreciate your thoughtful suggestion to highlight this contribution more clearly. In the revision, we will explicitly emphasize this as a novel experimental contribution.

Thank you again for your constructive comments, and we truly appreciate your feedback.

We appreciate Reviewer SFyk for recognizing the paper presentation, motivation, and in-depth theoretical exploration. We appreciate your time and constructive comments. We hope our clarifications below can address your concerns and raise your rating of the paper.

W1 and Q1: Problem definition, data and task descriptions.
Thank you for your comments. We would like to clarify that the problem definition is provided in Section 2.2, and the data and task description are detailed in Section 5.1 and Appendices F and G. The term "EEG snapshot" refers to each short segment (i.e. 1-second) of EEG data that is extracted by sliding a window over the continuous EEG signal, serving as the input to the RNN-based model (as illustrated in Figure 1). We will revise the final version to include a more detailed description and updated visualizations.

W2 and Q2: Novelty and components specifically tailored to this task.
Our time-then-graph architecture and explicit dynamic graph structure are specifically tailored to EEG tasks. That is, our model design is fundamentally guided by brain dynamics. The choice of the two-stream Mamba architecture is grounded in our time-then-graph design, which separates the temporal modeling of node and edge attributes for explicitly modeling brain dynamics. This architecture incorporates a selective mechanism that mirrors the neurobiological processes of selectively retaining relevant information and adaptively integrating new stimuli. In addition, we incorporate LapPE to preserve brain region specificity. We propose explicit dynamic graph structures for each EEG snapshot rather than using a single static graph, enabling it to capture transient connectivity changes of brain state, such as seizures. We hope that EvoBrain introduces a new architectural perspective to seizure research. Thank you for your insightful comments, we will highlight this discussion more clearly in the revised version and relate it to the relevant literature.

W3 and Q3: Recent models.
Thank you for your comments. We compared our method against LaBraM (ICLR'24), one of the most recent and powerful EEG foundation models. LaBraM is pretrained on large-scale datasets across multiple domains and possesses a larger number of parameters.
Moreover, we include two new baseline comparisons: EEGPT (NeurIPS’24) and AMAG (NeurIPS’24). EEGPT is a recent foundation model for EEG, while AMAG is a deep graph model focused on neural dynamics. Similar to LaBraM, EEGPT is pretrained on large-scale EEG datasets across multiple domains and contains 50 million parameters—approximately 277× larger than our EvoBrain model. On the 12-second seizure detection task, EEGPT and AMAG achieve AUROC scores of 80.4 and 81.7, respectively, while our EvoBrain model achieves 87.7. This further supports the effectiveness of the proposed method.
We will include all results in the final version.

W4 and Q4: Evaluation metrics.
In seizure tasks, seizure periods are typically much shorter than non-seizure periods, and our dataset reflects this class imbalance. In such cases, accuracy can be misleading, as a model that always predicts the majority class (i.e., non-seizure) may achieve high accuracy but provide little clinical value. In contrast, F1 score balances precision and recall, making it more appropriate when positive (seizure) instances are rare but important to detect. Similarly, AUC (Area Under the ROC Curve) evaluates the model’s ability to distinguish between classes across different thresholds and is less affected by class imbalance.
Our experimental settings are fully aligned with the reference work DCRNN [1] and GRAPHS4MER [2]. We will also include precision and recall metrics in the final version, following your suggestion.

• [1] Self-Supervised Graph Neural Networks for Improved Electroencephalographic Seizure Analysis, ICLR, 2022.
• [2] Modeling Multivariate Biosignals With Graph Neural Networks and Structured State Space Models. CHIL, 2023.

Limitations.
Thank you for your comments and for pointing out this aspect, and we will enhance the discussion of limitations in the final version of our paper. Below is our intended discussion:
In seizure prediction, where pre-ictal patterns are typically weaker and more spatially diffuse. While EvoBrain achieves the best performance among lightweight GNN-based models, LaBraM benefits from having approximately 30 times more parameters and pretraining on large-scale multiple EEG corpora, which enhances its ability to generalize under limited and noisy pre-seizure data conditions. While we focus our evaluation on the seizure task, which is particularly critical and life-threatening among various EEG applications, generalization to other tasks remains a limitation of our work.

We sincerely appreciate your recognition and are pleased to hear that our responses have addressed your concerns and contributed to the increased score. We thank you again for your kind suggestions and dedication to the review process.

We appreciate Reviewer qoYr for recognizing the solid mathematical foundation and acknowledging this as a strong manuscript. We appreciate your time and constructive comments. Below, please find our point-by‑point responses to your concerns and questions. We hope our clarifications below can address your concerns and raise your rating of the paper.

W1: LaBraM outperforms ours on seizure prediction task.
Thank you for your comments, and we will discuss this in the final version as a limitation. Below is our intended discussion:
In seizure detection, our method demonstrates superior performance due to its ability to highlight strong, localized network changes with high representational power. These changes are indicative of seizure onset and are effectively captured by our dynamic graph modeling.
However, in seizure prediction, where pre-ictal patterns are typically weaker and more spatially diffuse. While EvoBrain achieves the best performance among lightweight GNN-based models, LaBraM benefits from having approximately 30 times more parameters and pretraining on large-scale multiple EEG corpora, which enhances its ability to generalize under limited and noisy pre-seizure data conditions.

W2: Memory.
We appreciate the reviewer’s insightful comment regarding memory consumption and limitations. As suggested, we have measured the maximum GPU memory usage with a batch size of 1 for all GNN models.

While DCRNN and EvolveGCN are indeed more memory-efficient , our EvoBrain achieves over 10× faster compared to these baselines as shown in the figure 4. Among the three time-then-graph models, EvoBrain is the fastest and provides a well-balanced trade-off between memory usage and computational speed.
We will include these points in the final version of our paper.

We thank Reviewer Ma49 for recognizing our well-defined motivation, solid theoretical analysis, strong baselines, large realistic dataset experiments, and neuroscientific interpretability. We appreciate your time and constructive comments. Below, please find our point-by‑point responses to your concerns and questions, and we hope our clarifications below can address your concerns and raise your rating of the paper.

W1: The core contribution is primarily an integration and refinement of known components, rather than a fundamentally new paradigm.
Thank you for the discussion. Our contribution lies in the theoretical foundation for modeling brain dynamics, along with the proposal of novel seizure detection models based on this foundation.
As you mentioned, this work is to first investigate and provide theoretical analyses for EEG modeling using dynamic graph neural networks. From a technical perspective, while EvoBrain is not a fundamentally new paradigm, it is a first time-then-graph architecture that combines explicit dynamic modeling for seizure tasks, guided by our theoretical foundation, that effectively captures evolving brain network dynamics.

W2: The main difference here is in more rigorous theoretical justification and cleaner engineering.
We hope that moving from intuitive modeling to a theoretically guided and clearly structured approach can provide a foundation for future research on GNN-based seizure modeling.
Several recent works (e.g., GRAPHS4MER) learn explicit dynamic graphs internally; however, they begin with a temporally fixed input graph. In contrast, our method computes the graph structure initially and directly based on individual temporal snapshots (e.g., 1 second) from the beginning.

W3: Generalizability to more diverse EEG tasks.
Thank you for your discussion. Our results already demonstrate the effectiveness of our method in terms of cross-patient generalization. We would like to clarify that training and test sets are composed of entirely different patients, and we will explicitly state this in Line 268 of the revision.
We also appreciate your insightful comments. We will expand the discussion to include diverse EEG tasks and multi-modal fusion modeling in the Discussion Section.

W4: Some implementation choices (two-stream Mamba, LapPE) feel like plug-and-play improvements, and the overall novelty is incremental relative to the field’s direction. not big innovation.
Our model design is fundamentally guided by brain dynamics. The choice of the two-stream Mamba architecture is grounded in our time-then-graph design, which separates the temporal modeling of node and edge attributes for explicitly modeling dynamics. This architecture incorporates a selective mechanism that mirrors the neurobiological processes of selectively retaining relevant information and adaptively integrating new stimuli. In addition, we incorporate LapPE to preserve brain region specificity. We hope that EvoBrain introduces a new architectural perspective to seizure detection research.
We will further clarify these points in the subsequent vision of the paper.

Hello, Would greatly appreciate your response to the authors' rebuttal particularly at this critical point in time. Thanks Best