acceptodds
Under review as a conference paper at ICLR 2027

DANCE: Detect and Classify Events in EEG

Abstract

Event identification in continuous neural recordings is a critical task in neuroscience. Decoding in EEG is dominated by classifying windows aligned to known event onsets. However, while available in controlled experiments, such onsets are absent in continuous real-world monitoring. Here, we introduce DANCE, a deep learning pipeline that frames neural decoding as a set-prediction problem and jointly detects and classifies events directly from raw, unaligned signals. Evaluated separately on ten datasets curated from the literature with a wide variety of event types (ranging from milliseconds to minutes in duration), our model outperforms existing methods on a broad range of cognitive, clinical and BCI tasks. This single architecture establishes a new state of the art in the competitive task of seizure monitoring and matches the accuracy of onset-informed models for BCI tasks. Overall, our method marks a step towards end-to-end asynchronous neural decoding models.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neuroscience, deep learning, events, event detection, event classification

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_eAKjOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an end-to-end asynchronous EEG decoding framework named DANCE, which reformulates neural event detection and classification modeling as a set prediction problem. The authors conducted extensive evaluations across six task domains and ten public datasets, comprising a total of 1,154 participants and 2.35 million events.

Critical strengths
  1. Advance EEG decoding from classification based on idealized alignment windows to realistic scenarios involving asynchronous, continuous event detection.
  2. Successfully integrated the set prediction concept from computer vision (the DETR architecture) with Perceiver's flexible temporal-scale resampling mechanism.
  3. The evaluation is exceptionally robust, spanning 10 public datasets, 6 distinct task domains, and over 2.35 million events; it compares 7 different baselines and includes comprehensive ablation studies.
Critical weaknesses
  1. The core components of DANCE are built upon existing frameworks—specifically CNN, Perceiver, and DETR-style set prediction architectures; therefore, it is necessary to further elucidate the key methodological innovations tailored to the specific characteristics of EEG data, rather than merely presenting a combination of existing modules.
  2. The paper employs a combination of temporal IoU thresholds and class consistency to determine correct event detections. However, given the significant variations in event duration, event density, and localization difficulty across different tasks, the use of a single evaluation threshold may compromise cross-task comparability.
  3. Although the model was evaluated on multiple datasets, independent training configurations were used for each, and the authors noted that joint training has not yet yielded performance gains. Consequently, while the results demonstrate the architecture's adaptability to various tasks, they are insufficient to prove a unified capability for cross-task generalization.
Questions for authors
  1. The papers cover events ranging from short-duration speech-related events (approximately 0.08 seconds) to clinical events lasting tens of seconds or even minutes; there are significant differences across tasks regarding event duration, density, and boundary annotation characteristics. Applying a uniform IoU threshold of 0.5 may result in varying levels of actual strictness for the metrics across different tasks.
  2. If DANCE employs independently trained models for each dataset, it does not operate on the same comparative basis as large-scale, pre-trained EEG foundation models, which aim for universal representations and cross-task transferability. Consequently, one cannot conclude that this architecture possesses superior generality or is better suited for general-purpose neural decoding based solely on independent test results across datasets.
  3. For events spanning a vast range of durations (from 0.08-second phonemes to 67.7-second epileptic seizures), does a fixed sensitivity for box regression boundaries lead to imprecise boundary localization for long events or severe localization jitter for short events?

If the author addresses the questions above, I will raise my score.

2 comments

Thank you for your review. Here are our answers, in order.

1. IoU 0.5. We selected this metric as this is a demanding test. IoU divides the overlap by the length of the events, so a threshold of 0.5 asks for the same relative precision from an 80 ms phoneme and a 60 s seizure; a tolerance in seconds would be loose for seizures and out of reach for phonemes. F1eventF1_{\text{event}} decreases smoothly with the threshold, with no cliff at 0.5 (Fig. A.1A). Across all methods, DANCE has the highest mean F1eventF1_{\text{event}} at every threshold from 0.1 to 0.9, and its lead grows as the test gets stricter (1.8× at 0.1, 3.5× at 0.5, 4.7× at 0.9).

2. Generality. We agree that per-dataset training shows that the architecture adapts to each task, not that one model generalises across tasks. Our goal is to propose a recipe, not a pretrained model. On a new dataset, DANCE tests detection and classification from scratch, without pretraining, by setting two parameters (window and number of queries), with a rule for the window (App. A.3); we did this on ten datasets in six domains. It is also complementary to foundation models: our decoder turns pretrained REVE into an event detector (see our answer to Reviewer 9s8k). We have now made this explicit in the Discussion.

3. Boundaries. The window is set from the event length (8 s for phonemes, 256 s for seizures), and boundaries are predicted relative to the window, so precision scales with event duration. The median boundary error is about one latent token (23–129 ms) on six datasets, 1–2% of trial length in motor imagery, and 11% of seizure length. DANCE's relative error is lower than that of the best baseline on all ten datasets.

4. EEG-specific design. We agree that DANCE builds on existing modules. What is new is how each one addresses a property of EEG:

  • Low SNR and small datasets: a bare set loss reaches 0.326 F1eventF1_{\text{event}}; our dense and consistency losses raise it to 0.398 (Fig. 2A).
  • Event durations spanning four orders of magnitude: the Perceiver maps every window to 256 tokens, and a tokens-per-event rule sets the window; breaking this rule drops phoneme F1eventF1_{\text{event}} from 0.45 to 0.18 (App. A.3).
  • Varied montages (13 to 128 channels): sensor-position attention reads all of them.
  • No post-processing rules: previous event-based detectors for EEG, such as DOSED and EventNet, rely on hand-crafted rules (see our answer to Reviewer tm1p).
  • No event-level benchmark: we provide the metric and ten datasets.

Together, these let one recipe beat seizure-specific models on TUSZ without post-processing. Our goal here is to propose a reference architecture and to push the community to consider this problem, which is mostly overlooked in the literature.

Thank you for addressing my questions—best of luck to you! I will increase my rating.

Reviewer_tm1pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

DANCE, as a deep learning model for automatic EEG event detection, is proposed to identify event types and their start and end times without knowing the timing beforehand. This is important since in real-life applications such as seizure detection or BCI applications, the events happen at unknown times. The current gap is that existing methods miss important events, predict a label for the entire window, and require information about event timing. The Dance finds whether an event happened, what kind of event it is, and what its start and end times. Its architecture combines a CNN, a Perceiver, and a transformer decoder. The model is evaluated on ten datasets across six task domains.

Critical strengths

Critical Strengths:

  1. The model detects and classifies events directly from continuous EEG without known onsets, which is the setting real monitoring and asynchronous BCI require.
  2. One architecture covers events spanning several orders of magnitude in duration across six domains. Only window length, query count, and batch size are set per dataset, plus seizure-specific sampling on TUSZ.
  3. The ablations are informative. Dense readout scores 0.161, the set loss alone 0.326, and the full hybrid objective 0.398, with both steps significant. Loss weights are robust, and the window-size analysis gives a usable rule of thumb.
Critical weaknesses

Critical Weaknesses:

  1. The baseline comparisons may not fully explain DANCE's advantage. The baseline models identify events by grouping consecutive predictions of the same class. This means they cannot separate two events of the same class when they touch or overlap. This may partly explain why DANCE performs better, especially on speech datasets. Therefore, it is difficult to tell how much of the improvement comes from DANCE's better event detection approach versus limitations in how the baselines produce events. The authors should discuss this limitation and be more careful when concluding that pretrained EEG models do not transfer well to event detection.

  2. DANCE achieves 62.1% seizure F1 compared with Seizure Transformer's 58.0%, but this comparison uses sample-level scores and does not demonstrate better event-level seizure detection. For BCI, DANCE achieves 68.14% accuracy only on temporally matched predictions, which is still below the best onset-informed result of 77.82%. These are promising results, but the paper should be clearer about what the comparisons demonstrate.

  3. The paper uses EEG windows ranging from 8 to 256 seconds but does not clearly explain how events crossing window boundaries are handled or how predictions from different windows are combined. This could affect the reported event counts and false-alarm rates. Also, the reported inference latency only measures processing time and does not account for the time needed to collect the EEG data. Therefore, it is unclear how well the method would perform in real-time applications.

  4. On the ten datasets, DANCE is compared only with models that label each time point. Two EEG models the paper cites already predict events directly: DOSED (Chambon et al., 2019) and EventNet (Seeuws et al., 2024). Neither is tested on these datasets, and EventNet appears only as a literature score on TUSZ. Without them, it is unclear whether the gains come from DANCE’s design or simply from predicting events. This also limits the novelty, which is mainly applying DETR-style prediction and a Perceiver to EEG.

Questions for authors

Please mainly address the weakness points 2-4, to better demonstrate the performance improvement and clarify the novelty beyond mainly applying DETR-style prediction and a Perceiver to EEG.

4 comments

Answer (1/2): points 1–3.

Thank you for your thoughtful review.

1. Touching events. We thank the reviewer for pointing out this missing explanation, we should have explained it in the paper. A dense output gives one label per time step, so two touching events of the same class come out as a single segment. To avoid penalising the baselines for this, we scored every dense model against a version of the annotations in which such events are merged; a perfect dense model scores F1event=1F1_{\text{event}} = 1 on it. DANCE was scored against the original, separate events, which is harder, and still leads on every dataset. On motor imagery, where no events touch, the two versions are identical and DANCE leads by 0.11 and 0.32. For our point on the foundation-model claim: we will reword it as "a dense readout of a pretrained backbone does not give events, while our decoder does".

2a. Seizures at the event level. DANCE also leads at the event level (Table A.1): F1eventF1_{\text{event}} 0.21 against 0.02 for the best baseline. At the false-alarm rates used in clinical monitoring, with each baseline given its best threshold on the test set, DANCE detects 28, 33, 39 and 46% of seizures at 1, 2, 5 and 10 false alarms per day, against 21, 22, 29 and 34% for REVE, the best baseline. Fig. 4A uses sample-level F1 because published TUSZ results report it. We report a sample-based metric for this application as this is the standard in seizure detection community.

2b. BCI. We acknowledge that our wording was unclear. "Temporally matched" is not a subset: DANCE finds 98.6% of the trials, so the 68.1% accuracy covers almost all of them. DANCE solves two tasks, finding each trial and classifying it, while onset-informed pipelines are given the trial window and only classify. To make the comparison fair, we need to disentangle both tasks, by selecting events that are correctly localized, as the classification metric would not be comparable otherwise. It still beats their average (61.3%). We will rewrite this in the paper.

3a. Window borders. We thank the reviewer for this question. DANCE predicts each event's start and end rather than its centre and duration, so no prediction depends on a centre outside the window. Test windows are consecutive, and an event cut by a border is left out of that window's score (13% of events). Scoring whole recordings against all events, including the cut ones, gives F1eventF1_{\text{event}} 0.38 instead of 0.40; false alarms per hour change by less than 5%, and the best baseline (CBraMod) stays at 0.10. We will report this in the paper.

3b. Real time. Processing one window takes 13 ms on a single GPU. With a buffer that moves every 0.25 s, DANCE reports an event at most 0.46 s after it ends (the worst case, for seizures). Onset-informed BCIs also wait for the end of the trial before they can classify it. We added this precision in the paper.

Point 4 follows in the next comment.

Answer (2/2): point 4.

4. DOSED and EventNet. We thank the reviewer for pointing out these two works. They differ from DANCE as follows:

  • One pass for all classes: EventNet detects one event type at a time (its authors merge muscle and chewing artefacts "into a single artefact class"). DANCE detects and classifies all classes jointly.
  • No hand-set priors: DOSED needs default event durations for each event type, and EventNet a maximum duration and a backbone tailored to each dataset. DANCE keeps one architecture for events from 80 ms to minutes and only sets the window size and the number of queries.
  • No post-processing: both need peak picking or non-maximum suppression with tuned thresholds. DANCE outputs events directly.
  • Six domains instead of one: DOSED targets sleep events, EventNet artefacts and seizures.
  • A training objective suited to EEG: a bare set loss is not enough; with our backbone it reaches 0.326, and the full DANCE, with its dense and consistency losses, 0.398 (Fig. 2A).

We trained EventNet on five of our datasets, with one output per class and the same tuning for both models. After picking reasonable values, DANCE is still ahead on average (F1eventF1_{\text{event}} 0.414 vs 0.354), with the largest gaps on motor imagery (0.22 and 0.11), where EventNet labels trials near chance. Moving EventNet's peak threshold by 0.1 drops its score on BI2014b from 0.43 to 0.05. On TUSZ, its own task, it is below DANCE in Fig. 4A (48.3% vs 62.1%). Above all, DANCE does not rely on any post-processing rule, which we see as its main advantage.

We trained DOSED with the authors' code, under the same protocol and tuning as EventNet. DOSED requires many hand-set choices. It is an anchor-based detector: you pick default event durations (anchors) for each dataset and a suppression threshold. On motor imagery, where all trials have the same duration, DOSED ties DANCE on BNCI2014 (0.429 vs 0.414) and is clearly below it on BNCI2015 (0.406 vs 0.624). On seizures, which last from seconds to minutes, DANCE detects 2 to 5 times more seizures at 1 to 10 false alarms per day (28–46% vs 6–25%).

Many thanks for addressing my comments. It certainly helps to better understand the paper settings.

Thank you for your reply, we are glad the answers clarified the settings. If any concern still stands in the way of a more positive assessment, we would be happy to address it before the discussion closes.

Reviewer_9s8kOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper, DANCE is introduced as a novel pipeline for jointly detecting and classifying events in raw, unaligned EEG signals. This addresses the still dominant reliance on aligned windows and expert knowledge and is a step towards online BCI models. They test their model on ten existing datasets spanning a range of event types and durations. DANCE marginally outperforms existing models on the standard sample-based F1 score, while strongly outperforming them on their new event-based metric which penalizes segmented event detection.

Critical strengths
  1. This study addresses an important problem in EEG decoding and reformulates it as set prediction. The method manages to jointly predict event time and duration as well as event type with decent performance gains over existing methods.
  2. The method is tested across diverse datasets and event duration and type which points towards the generalizability of the model and its possibility for wide application. Further, the ablation studies show the importance of the model components.
  3. The event-based metric is valuable and logically constructed. It is a useful evaluation measure for future studies. While the performance gains of DANCE are most evident on this metric, I see its advantage of actually measuring the output we desire, namely free detection of full events and classifying their type without requiring post processing or combination of split detections.
  4. The paper is very well structured and easy to follow. Substantial additional interesting information is provided in the appendix.
Critical weaknesses
  1. The performance gains are only small on the sample-based metric and concentrate on their own event-based matric, which matches DANCE's training objective. As the comparison models are not specifically made to perform well on this metric, this is an unfair comparison.
  2. The overall absolute performance is modest, with average event F1 score of 0.39, which indicates that many events are missed or wrongly classified, which limits applicability.
Questions for authors

The largest performance gain in the F1 Event score comes from adding the decoder (Figure 2A). The comparison models do not have a decoder that is designated for this task (basically maximizing F1 event). How do these models compare when adding the DANCE-style decoder?

2 comments

Thank you for the positive review.

Baselines with our decoder. We added our Perceiver and decoder on top of pretrained REVE, the largest foundation model we test:

F1eventF1_{\text{event}}BI2014bBNCI2014BroderickBNCI2015TUAR*
REVE + our decoder.366 ± .014.294 ± .006.444 ± .004.320 ± .018.135 ± .008
DANCE.392 ± .013.398 ± .023.446 ± .004.573 ± .041.138 ± .007

Our decoder turns REVE into an event detector (up to 10× higher F1eventF1_{\text{event}}), which supports the main claim of the paper that set-based methods improve event detection on EEG. DANCE, smaller and trained from scratch, comes first on all five datasets. We will add the other datasets to the final version.

*REVE was pretrained on the TUH corpus, which contains TUAR.

Fairness. We agree with the reviewer that F1eventF1_{\text{event}} matches what DANCE is trained for. We introduced it because DANCE is designed to detect and classify whole events, and existing metrics on classical tasks do not measure this. To keep the comparison fair, we also report the baselines' own metric, F1sampleF1_{\text{sample}}, on which DANCE is tied with the best of them (Fig. 2A), and we gave the dense baselines additional post-processing (App. B.2). We have now also compared DANCE with models built for events: it still leads REVE with our decoder (above) and EventNet, trained on five of our datasets with the same tuning as DANCE (mean F1eventF1_{\text{event}} 0.414 vs 0.354; see our answer to Reviewer tm1p). We believe this is a fair assessment as only few existing methods consider this problem directly but we are happy to report additional metrics/baseline if some are missing.

Absolute performance. Absolute scores are low partly because F1eventF1_{\text{event}} is demanding: right class, at least half overlap with the true event, one prediction per event, unseen subjects and no onsets. To complement this score, we also report downstream performances where an application uses the output:

  • BCI (BNCI2014): 98.6% of trials found; 68.1% classified correctly without onsets, above the average onset-informed pipeline (61.3%). Direct comparison with top performing method seems unfair as the onset information bias the comparison.
  • Seizures (TUSZ): state-of-the-art seizure F1 (Fig. 4A). At 1 to 10 false alarms per day, DANCE detects 28–46% of seizures, against 21–34% for the best baseline.

We agree that there is still improvement needed to obtain a high-performing detector and classifier. For example, F1eventF1_{\text{event}} stays below 0.15 on artefacts (TUAR). The event-level metric makes this gap visible, and we hope it will help the field measure progress.

Thank you for addressing my comments. I think the paper is interesting and I still believe the F1_event metric together with DANCE is valuable. I will leave my score for now and wait for the AC discussion.

Comments on the paper

1 comment

We thank the reviewers for their comments on the paper. Overall, they agree that the problem is relevant, and needs event-level metrics. They also note the breadth of the evaluation (ten datasets, six domains).

The open questions were about additional comparisons, that we added to the paper:

  • Foundation-model baselines with our decoder (9s8k);
  • EventNet and DOSED trained and tested (tm1p);
  • Seizure detection at clinical false-alarm rates (tm1p);
  • Whole-recording scoring, including events cut by window borders (tm1p);
  • All methods across IoU thresholds, and boundary error across event durations (eAKj).

DANCE stays ahead in each of them, as described in individual answers. We are ready to run further analyses during the discussion if some comparisons need to be expanded. We have also submitted a revised version of the PDF incorporating our new explanations and additional experiments.

We believe DANCE is a useful addition to the literature as it proposes one recipe to detect events without onsets or post-processing, across six domains with heterogeneous settings. It greatly improves the event-level score of the best baselines.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.