acceptodds
Under review as a conference paper at ICLR 2027

NeuroStream: A video-inspired causal spatiotemporal encoder for streaming EEG

13 views

Abstract

This paper presents a hybrid-architecture foundation model for multichannel EEG that models spatiotemporal features separately. EEG foundation models have developed rapidly in recent years, and most adopt ViT-like paradigms: they treat continuous EEG signals as static patch grids, divide the signal into patches (tokens), flatten them into a sequence, and feed the sequence into a Transformer. This approach requires users to carefully set the temporal window size for each task to obtain the best performance, and the window size cannot be changed while the model is operating. Unlike static images, EEG is structurally more similar to video, with distinct temporal and spatial dimensions. Inspired by work on streaming video understanding, we regard continuous EEG signals as short frames arriving over time and process their spatial and temporal dimensions differently: within each frame, a Transformer models spatial relationships among electrode channels, while a causal Mamba state-space module maintains cross-frame temporal states for each channel. When a new frame arrives, the model only needs to perform incremental computation. It can accommodate EEG signals of different lengths and, through learnable state-transition parameters, form task-related effective context ranges, thereby reducing reliance on preset window lengths. We construct a self-supervised pretraining task on large-scale EEG datasets: next-frame amplitude-spectrum prediction. Given historical EEG frames, the model predicts the amplitude spectrum of each channel in the next frame, frame by frame, enabling it to learn cross-frame frequency-domain changes in continuous EEG signals. We conduct experiments on seizure detection, sleep staging, and emotion recognition. The results show that the proposed pretrained model transfers effectively to different downstream tasks, achieving a **balanced accuracy above 95%** and an **AUROC of 0.99**, particularly on the cross-subject CHB-MIT seizure-monitoring task.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Electroencephalography (EEG), EEG Foundation Models, Self-Supervised Learning, State Space Models, Spatiotemporal Modeling

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kfVoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The manuscript presents NeuroStream, a streaming EEG foundation model that combines within-frame spatial Transformer attention with channel-wise causal Mamba state-space modeling. It also introduces next-frame amplitude-spectrum prediction for self-supervised pretraining and evaluates the model across five EEG downstream tasks. The problem of processing continuous EEG signals is relevant, and the proposed framework demonstrates promising performance on long-duration tasks. However, the methodological novelty is limited, as the core framework primarily integrates established spatial attention, state-space modeling, and frequency-domain pretraining techniques. The distinct technical contribution of this particular integration is not sufficiently established.

Critical strengths
  • The paper addresses a meaningful practical problem: modeling continuous EEG signals without relying on fixed temporal windows or repeatedly processing historical observations.
  • The separation of spatial attention and temporal state-space modeling provides an intuitive and computationally efficient architecture for streaming EEG analysis.
Critical weaknesses
  • The proposed architecture is not sufficiently differentiated from existing spatiotemporal modeling approaches, such as CBraMod. Replacing temporal attention with a standard Mamba-based state-space module enables causal streaming but does not by itself establish a substantial methodological contribution.
  • The claimed advantages of bounded memory, variable-length inference, and adaptive temporal context largely follow from established properties of causal state-space models. The manuscript does not introduce a clearly distinct mechanism beyond their application to EEG.
  • The proposed next-frame amplitude-spectrum prediction combines established autoregressive prediction and frequency-domain pretraining strategies. Its methodological novelty and specific advantages over existing EEG pretraining objectives remain insufficiently demonstrated.
  • Although the experiments show promising performance, the ablations do not adequately isolate the benefits of the proposed architectural integration or pretraining objective. Moreover, different temporal input lengths between streaming and non-streaming baselines complicate direct performance comparisons.
Questions for authors
  • Beyond replacing temporal attention with Mamba, what specific methodological contribution distinguishes NeuroStream from existing spatiotemporal EEG architectures such as CBraMod? Is there an EEG-specific mechanism that cannot be achieved through a straightforward combination of existing components?
Reviewer_5m19OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes the NeuroStream streaming EEG basic model to address the streaming modeling issues of existing EEG basic models that rely on fixed time windows and are difficult to efficiently process continuously variable-length long-duration EEG signals. While effectively enhancing the performance of long-term clinical EEG modeling, it still remains competitive in short-term BCI tasks. Based on a large number of experimental results, its effectiveness has been fully verified.

Critical strengths

This paper proposes a hybrid streaming EEG base model (NeuroStream), which combines intra-frame spatial Transformer attention with cross-frame, channel-level causal Mamba state space update and uses the amplitude spectrum prediction of the next frame as the pre-training objective. The evaluation covers multiple paradigms and multiple datasets and validates the montage transfer under the leakage-safe per-subject division; the authors also provide a comparison of matching windows for each baseline and various model scales and clearly report that the model lags behind the strongest supervised baseline in short BCI tasks, making the empirical conclusion more reliable. The engineering motivation of this paper is clear and timely, the experiments are comprehensive enough, and the matching windows and ablation results clearly reflect the advantages of the proposed streaming design.

Critical weaknesses
  1. The core positioning of this article has significant overlap with existing research, and the increment in novelty needs to be further clarified.
  2. The comparisons relied upon in the main results were not conducted under the same window settings, and the key details of the evaluation protocol are insufficiently described.
Questions for authors
  1. The core positioning of this article has significant overlap with existing research, and the increment in novelty needs to be further clarified.
  2. The comparisons relied upon in the main results were not conducted under the same window settings, and the key details of the evaluation protocol are insufficiently described.
Reviewer_V4CrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes NeuroStream, a causal EEG encoder that treats a recording as a stream of 1-s frames. Within each frame, a Transformer attends across electrode channels. Across frames, a per-channel Mamba-3 recurrence carries a fixed-size state, so each new frame is encoded incrementally and the total cost grows linearly with recording length for a fixed number of channels. The encoder is pretrained on four TUH subsets by predicting each channel’s amplitude spectrum in the next frame, and is then fine-tuned on five tasks. The strongest result is on CHB-MIT seizure detection, with a balanced accuracy (BalAcc) of 0.955 for 300-s inputs and 0.959 for 600-s inputs. Results on the other tasks are weaker. On Sleep-EDFx, NeuroStream achieves the highest BalAcc but clearly lower weighted F1 and Cohen’s kappa than BIOT and CBraMod. On BNCI2014004 and BNCI2014009, it trails the best supervised baselines.

Critical strengths
  1. The architecture is well matched to the streaming problem it targets. Channel interactions are handled within each frame, and temporal context is carried in a recurrent state, so past frames never need to be re-encoded. Next-frame spectrum prediction is a natural pretraining objective for such a causal encoder. The FLOP and memory measurements in Appendix F (Table 13, with slopes close to 1) support the linear-cost claim.

  2. Pretraining has a large and consistent effect. Across all head sizes tested, pretrained initialization raises CHB-MIT BalAcc from about 0.75 to about 0.95. In the matched 1-s CHB-MIT setting (Table 11), NeuroStream still achieves the highest BalAcc (0.878), although its AUROC and AP are comparable to those of ShallowNet and EEGNet.

Critical weaknesses
  1. The architectural novelty is limited, as each component is drawn from prior work. The within-frame channel attention follows the divided space-time design of TimeSformer, which CBraMod has already adapted to EEG. The temporal module is a standard Mamba-3 layer applied to each channel. The pretraining target follows earlier frequency-domain objectives, and the next-step prediction follows Neuro-GPT. What is new is the combination of these components for causal, frame-by-frame inference. This combination is reasonable, but it is mainly an engineering choice, so the paper needs to show that it outperforms simpler alternatives. The paper also does not compare against existing Mamba-based EEG models such as EEGMamba and FEMBA.

  2. The experiments do not disentangle the contributions of the temporal module, pretraining, model size, and access to history. In Tables 2 and 3, NeuroStream receives 300–600 s (CHB-MIT) or 1800–3600 s (Sleep-EDFx) of signal, whereas the pretrained baselines see independent 1-s or 30-s windows. When the input setting is matched (Tables 11 and 12), however, most of NeuroStream’s advantage disappears. The streaming LSTM and Mamba-3 baselines do not resolve this issue. They are much smaller than NeuroStream (0.15–0.87M vs. 5.5–7.7M parameters), and the paper does not state whether they were pretrained. These baselines are near chance on CHB-MIT, yet replacing Mamba with an LSTM inside NeuroStream costs only about 12 points (Figure 3), which points to a difference in setup rather than in architecture. The ablations in Figure 3 likewise do not report their pretraining status or parameter counts.

  3. There is no direct evidence that the model uses long-range context, and the sleep-staging results show a trade-off. Extending the input from 300 to 600 s, or from 1800 to 3600 s, shows that the model can run on longer inputs. However, the gains are small (e.g., Sleep-EDFx BalAcc increases by less than 0.01, from 0.750 to 0.759), and these experiments do not reveal how much of the history actually influences the predictions. On Sleep-EDFx (Table 3), NeuroStream achieves the highest BalAcc, but its weighted F1 (0.646) and kappa (0.525) are well below those of BIOT (0.775 / 0.660) and CBraMod (0.754 / 0.632). Higher BalAcc combined with lower weighted F1 and kappa suggests that performance on the minority sleep stages improves at the expense of the majority stages. Together with the matched 30-s results in Table 12, this does not support the conclusion that the model “performs well” on both short and long tasks.

Questions for authors
  1. As noted above, the standalone LSTM and Mamba-3 baselines are close to chance on CHB-MIT, yet replacing Mamba with an LSTM inside NeuroStream costs only about 12 points (Figure 3). This suggests that the standalone baselines may have been set up differently from NeuroStream. Were they pretrained with the same data, objective, and training budget? What tokenization, parameter counts, and optimizer settings did they use? Their sizes also change from 0.87M/0.66M in Table 2 to 0.15M/0.30M in Table 3. What accounts for this difference?

  2. How is the recurrent state handled at segment, recording, and subject boundaries? Do the 600-s and 3600-s evaluations cover the same time points as the 300-s and 1800-s evaluations? The paper states that the model learns a task-relevant effective context, but no experiment measures how much history it actually uses. I would like to see how performance changes when the state is reset at shorter intervals, for example every 10, 30, or 60 s on CHB-MIT.

  3. Table 2 compares NeuroStream on 300/600-s sequences against non-streaming baselines on 1-s windows. Some of these baselines change substantially between Table 2 and Table 11, even though they receive 1-s windows in both (e.g., EEGNet AUROC 0.744 vs. 0.959 and BalAcc 0.657 vs. 0.803). What differs between the two setups? For example, do they differ in the training data after undersampling, the class ratio, or threshold selection? Under the matched 1-s setting, NeuroStream’s BalAcc lead shrinks from about 12 points to about 6, and it no longer holds a clear advantage in AUROC or AP. On Sleep-EDFx with 30-s inputs (Table 12), it falls below CBraMod on all three metrics. In the long-input setting, why does NeuroStream achieve the highest BalAcc while its weighted F1 and kappa remain well below those of BIOT and CBraMod?

Comments on the paper

1 comment
Area_Chair_rkrhOpenReview ↗

Reviewer Feedback and Final Discussion

Dear Reviewers,

The authors have submitted their rebuttals addressing your review comments, and the reviewer–author discussion period is coming to an end.

Could you please review the authors' responses, provide your feedback (if you have not), and actively engage in the final discussion before the deadline?

Thank you for your time and contributions.

Best regards,

The AC

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.