acceptodds
Under review as a conference paper at ICLR 2027

TRACE: Episode-Local Transition Evidence for Zero-Shot Seizure Onset Localization in SEEG

Abstract

In SEEG-based automated seizure analysis, accurate onset localization is important both for modeling seizure evolution and for supporting downstream clinical analysis. Existing methods face two key challenges: (1) detecting seizures reliably and localizing onset without priors such as patient identity, seizure type, or prior knowledge of seizure presence; and (2) generalizing across patients and centers despite substantial variation in background activity, seizure morphology, and acquisition conditions, without target-domain annotations, retraining, or parameter tuning. We propose TRACE, a framework based on episode-local transition evidence. TRACE encodes background-contrast features from each input clip into a shared channel–time evidence field that captures seizure-related changes relative to the episode-local background. A clip-level branch then summarizes that shared representation for seizure detection. When a clip is predicted as seizure-positive, the onset-localization branch selects high-evidence channels dynamically at each time point and constructs a global evidence trajectory by cross-channel pooling. A self-calibrated decoder reads that trajectory, identifies sustained seizure evidence, and backtracks to estimate the onset time. We evaluate our approach on SWEC-ETHZ and HUP under within-domain cross-patient settings. External zero-shot evaluation uses the patient-disjoint Pennsieve-530 cohort. On SWEC-ETHZ and HUP, our approach achieves oracle onset MAEs of 4.52 s and 4.92 s, reducing error by 37.7% and 34.0% relative to the strongest baseline in each cohort. On the spontaneous seizures of Pennsieve-530 (44 seizure and 132 non-seizure clips), our approach achieves a detection AUROC of 0.913 and an F1 score of 0.650, compared with 0.829 and 0.537 for the strongest baselines, while achieving an oracle onset MAE of 4.80 s, a 46.0% reduction relative to the strongest onset-localization baseline (8.89 s). These results show that constructing episode-local transition evidence and sparsely aggregating it provides an effective basis for cross-patient SEEG onset localization without target-domain adaptation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
seizure onset localization, zero-shot generalization, stereo-EEG, intracranial EEG, episode-local reference, domain shift, epilepsy

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_327POpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes TRACE, a framework for SEEG seizure detection and onset-time estimation without target-domain tuning. Each channel of a 60 s clip is described by 23 hand-crafted features measuring change relative to the clip's first 8 s, which serve as the episode's background. An XGBoost ensemble maps these features to a channel × time evidence field. A detection branch summarizes the field to decide whether a seizure is present, and an onset branch pools the top channels at each time point and backtracks from sustained high evidence to estimate the onset. Trained on SWEC-ETHZ and HUP and applied zero-shot to Pennsieve-530, TRACE outperforms the strongest of eight baselines in both onset error (oracle MAE 4.80 s vs. 8.89 s) and detection AUROC (0.913 vs. 0.829). Normalization control experiments suggest that the episode-local reference is the main source of this gain.

Critical strengths
  1. The framework is simple and computationally inexpensive, which keeps the cost of practical deployment low.
  2. The approach is zero-shot, with no target-domain tuning. This is a meaningful direction for the medical domain, where heterogeneity across patients and institutions is substantial.
  3. The results are analyzed through a variety of control experiments and multiple metrics. These include comparisons against alternative normalization schemes and experiments that change the source of the reference.
Critical weaknesses
  1. The core assumption is guaranteed artificially by the data construction. The method assumes that the first 8 s of every clip are always pre-ictal. This holds only because clips are constructed with the onset placed between 10 and 40 s. In realistic settings, where a clip may begin during or immediately after a seizure, this assumption is unlikely to hold.
  2. Novelty is limited. The method combines hand-crafted features, conventional ML (XGBoost), and a rule-based decoder, and the extracted features themselves offer little novelty.
  3. Seizure specificity is not evaluated. Background-contrast features may also respond to non-seizure events such as artifacts, interictal events, and sleep/arousal transitions. However, no evaluation is provided on negatives composed of such events.
  4. The treatment of heterogeneity is not sufficiently validated. It is unclear whether normalizing amplitude and spectral scale alone is sufficient to address heterogeneity. Subgroup analyses by seizure type, electrode type, and institution are needed.
  5. The external validation set is small. It contains only 44 seizures from 22 patients, which makes it difficult to assess the reliability of the zero-shot performance. In addition, this cohort had already been evaluated with an earlier configuration, and 19 of its 29 patients come from the same institution as a training cohort (HUP).
Questions for authors
  1. How does the method perform when the first 8 s contain a seizure or artifacts? How does it perform when applied in a sliding manner to continuous recordings, without the constraint on onset position?
  2. What is the false positive rate on negative clips composed of artifacts, interictal events, or state transitions?
  3. How were the weights of the threshold objective (Eq. 12; 5/3/2/2/3) chosen? Also, since the AUROC term does not vary with the threshold, does it not have any effect on the optimization?
  4. The rationale behind several design choices is unclear, including the 8 s reference, the three-model ensemble, and the 20/80 percentiles. Could the authors explain the specific reasoning behind each choice and provide sensitivity analyses?
Reviewer_S5viOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents TRACE, a carefully crafted framework for seizure onset detection and localization. Trace utilize episode-local transition evidence as features, together with XGBoost and two detection branch to achieve the best localization precision on two dev cohort and one held-out Pennsive dataset.

Critical strengths
  1. Important problem formulation: The paper is trying to address an important clinical problem: seizure onset localization. And the paper also try to focus on the generalizability across unseen dataset, which is very important to real clinical usage.

  2. The proposed pipeline is simple and human-interpretable. The idea of using background as reference, together with aggregation, two branches, etc are well motivated. It is also very interpretable compare to black-box ML models.

  3. The evaluation and experiment records is very transparent. I appreciate that the paper explicitly discussion the details and limitations of its evaluation setting and provide the iteration details in the appendix.

Critical weaknesses
  1. The Evaluation setting is not realistic. In abstract, the paper is saying that we want "detecting seizures reliably and localizing onset without priors". But in the paper, all positive clips are rubric crafted using ground-truth seizure onset time. It also make sure the onset is between 10-40s. Such crafting makes the first 8 seconds to be always pre-onset and can be used as background reference. In reality, we will not have such perfect setups: we cannot make any assumption about when the seizure onset happens (or if it happens at all). I would like to reframe the problem that the paper is trying to solve: Given a 60-second recording with a guaranteed pre-onset 8s starter, can the model estimate the seizure onset time that is guaranteed in the middle of clips? Such framing seems to be too narrow.

  2. The baseline is too weak. The best onset localization baseline (ViT-SOD) IS REPORTING mae OF 8.89, which is same as "constant 25 s guess scores 8.89 s on Pennsieve-530". This is basically saying that the best baseline model is no better than a trivial constant predictor. This makes the baseline model selection questionable. Can we use stronger baseline model, such as foundation models like Brant or PopT [1,2]?

Also, some machine learning baselines are data-hungry, and training with only 149 positive clips might be too few (The paper also state that the training is "bounded by compute").

In addition, the design of Trace explicitly tell it that "first 8s is the background reference", while the rest of the baseline has to learn implicitly.

  1. The held-out validation is not perfectly held-out. If I am checking correctly, more than half of the patient from Pennsive is coming from the same institution as HUP, so it is not a true held-institution out experiment. Also in appendix, it is stated that Pennsive has already been evaluated and analyzed using a earlier version of the framework, therefore it is hard to consider Pennsive as a truely held-out validation to evaluate the zero-shot capability of the framework.

[1] Geeling Chau, Christopher Wang, Sabera Talukder, Vighnesh Subramaniam, Saraswati Soedarmadji, Yisong Yue, Boris Katz, & Andrei Barbu. (2025). Population Transformer: Learning Population-level Representations of Neural Activity.

[2] Daoze Zhang, Zhizhang Yuan, Yang Yang, Junru Chen, Jingjing Wang, & Yafeng Li (2023). Brant: Foundation Model for Intracranial Neural Signal. In Thirty-seventh Conference on Neural Information Processing Systems.

Questions for authors
  1. Can we use stronger baseline model, such as foundation models like Brant or PopT [1,2]?

  2. Since in the appendix, it is saying that Pennsieve had already been evaluated using an earlier version of the framework, how were the final features selected, and what criteria determined the change from the earlier 13-feature subset?

  3. Can the authors provide learning curves for the deep learning baselines to make sure it is converging?

  4. Is it possible to evaluate using sliding windows on real continuous recordings and report onset error together with false positives?

[1] Geeling Chau, Christopher Wang, Sabera Talukder, Vighnesh Subramaniam, Saraswati Soedarmadji, Yisong Yue, Boris Katz, & Andrei Barbu. (2025). Population Transformer: Learning Population-level Representations of Neural Activity.

[2] Daoze Zhang, Zhizhang Yuan, Yang Yang, Junru Chen, Jingjing Wang, & Yafeng Li (2023). Brant: Foundation Model for Intracranial Neural Signal. In Thirty-seventh Conference on Neural Information Processing Systems.

Reviewer_ktD5OpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TRACE, a framework for patient-independent seizure detection and onset localization in stereoelectroencephalography (SEEG). The central idea is to represent seizure onset as a transition from an episode's own background rather than attempting to identify absolute seizure patterns across heterogeneous patients and recording systems.

TRACE first constructs background-referenced descriptors using the initial eight seconds of each 60-second recording. Three XGBoost classifiers transform these descriptors into a channel–time seizure evidence field. A clip-level classifier performs seizure detection, while the localization branch selects high-evidence channels, aggregates their temporal evidence, and estimates the seizure onset through self-calibrated thresholding and anchored backtracking.

The framework is evaluated on SWEC-ETHZ and HUP using patient-disjoint splits, followed by zero-shot cross-cohort evaluation on Pennsieve-530. The external evaluation includes 44 spontaneous seizures and 132 non-seizure clips. TRACE reports an oracle onset MAE of 4.80 seconds, compared with 8.89 seconds for the strongest baseline, alongside improvements in detection AUROC and F1. The authors also provide controlled referencing experiments and component ablations to investigate the source of these improvements.

Critical strengths

S1. Strong and relatively well-controlled evidence for the benefit of episode-local referencing.

The most convincing contribution is the careful evaluation of the referencing mechanism.

In the external Pennsieve-530 experiment, removing the episode-local reference increases oracle MAE from 4.80 to 7.54 seconds. The paper further evaluates alternative normalization approaches, changes in reference identity, reference positioning, and partial contamination with ictal activity.

The sequence of reference-identity interventions is particularly informative: onset MAE increases from 4.80 seconds with the original episode reference to 6.70 seconds using another seizure of the same patient, 8.04 seconds using interictal activity from that patient, 9.76 seconds using another patient, and 12.28 seconds using another site.

These experiments provide meaningful evidence that the benefit arises from the temporal and episode-specific nature of the reference rather than normalization alone.

S2. Meaningful external evaluation with a clear distinction between oracle and deployed performance.

The authors evaluate TRACE on an external, patient-disjoint cohort without target-domain adaptation and compare it against eight reimplemented baselines.

The separation of oracle onset MAE from deployed MAE is especially important. The former measures localization even when detection fails, whereas the latter explicitly penalizes missed seizures. This helps prevent the localization results from being interpreted independently of the detection system.

The paper additionally documents baseline fallback behavior and reports a matched-fallback sensitivity analysis, improving transparency in comparisons where certain baseline methods fail to return valid onset estimates.

Critical weaknesses

Critical Weaknesses W1. The experimental protocol guarantees the availability of a clean pre-onset reference and restricts seizure onset to a predefined portion of each clip.

The central concern is that seizure-positive clips are constructed with their annotated onset uniformly positioned between 10 and 40 seconds of a 60-second recording. Consequently, the initial eight-second segment used by TRACE as its episode-local reference is guaranteed to be seizure-free.

This is a strong methodological assumption. In realistic continuous monitoring, seizure onset may occur immediately after a recording begins, multiple seizures may occur within a short interval, or an appropriate clean baseline may not be available at a fixed location.

The experimental design therefore eliminates one of the principal practical difficulties of identifying a valid baseline without knowing the seizure onset in advance.

Although the authors investigate reference contamination and alternative reference positions, these analyses do not fully replace evaluation in continuous, unconstrained recordings where the system must determine whether a suitable reference exists.

Furthermore, the constrained onset interval itself supplies prior information that may substantially simplify localization. The constant 25-second baseline has an MAE of 8.89 seconds on Pennsieve-530, illustrating how informative the recording construction can be even without any EEG evidence.

This issue is consequential because the paper emphasizes zero-shot onset localization without seizure-presence or patient-specific priors. The current protocol supports that claim within a restricted clip-based setting, but does not establish comparable performance in unrestricted continuous SEEG.

W2. The evidence for independent external generalization is limited by cohort size, institutional overlap, and prior exposure to the external evaluation cohort.

The headline zero-shot results are obtained from 44 spontaneous seizures belonging to 22 patients. Although the external patients are disjoint from the development cohorts, the authors report that 19 of the 29 external patients were recorded at the same institution as the HUP development data.

This means that the experiment establishes cross-cohort and cross-patient generalization but cannot be interpreted as a completely independent cross-institutional evaluation.

The authors also disclose that an earlier model configuration had already been evaluated on Pennsieve-530 before the final TRACE configuration was frozen. Even without direct fine-tuning on the external labels, previous evaluation can influence research decisions and weaken the independence of a benchmark as a final confirmatory test.

The limited sample size further affects the strength of the comparative results. Appendix F indicates that, after Bonferroni correction, the improvement over the strongest localization baseline, ViT-SOD, does not remain statistically significant.

The paper also reports that the deployed-MAE improvement attributable to episode-local referencing has a confidence interval overlapping zero on Pennsieve-530.

These qualifications do not negate the promising results, but they materially weaken the strength of the paper's generalization and superiority claims.

Questions for authors

Q1. How does TRACE perform when seizure onsets are not restricted to the 10–40-second interval and the first eight seconds are not guaranteed to contain clean background activity? Can the authors evaluate randomly sampled continuous-recording windows without onset-conditioned placement?

Q2. How robust is TRACE to sequential seizures, seizures occurring near recording boundaries, and recordings without a clearly identifiable preictal baseline? Is there a mechanism for detecting when the assumed episode-local reference is invalid?

Q3. Can the authors report results separately for Pennsieve-530 patients recorded at institutions absent from the development cohorts? This would help distinguish cross-patient generalization from genuinely independent cross-site transfer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.