acceptodds
Under review as a conference paper at ICLR 2027

CHRONOPROTO: WHAT A SINGLE POSITIVE EPISODE CAN AND CANNOT PERSONALIZE

Abstract

When a seizure forecaster is personalized after a patient’s first usable seizure, the only labeled target data are typically the preceding preictal EEG. This episode locates the patient’s preictal windows, while ranking later windows also requires a reference for the interictal background. We formalize a positive-only protocol that releases this first preictal episode and keeps target background out of fitting. Under this protocol, ChronoProto contrasts a first-episode prototype with a fixed reference in a frozen source representation. We split the pairwise ranking error into prototype error and reference mismatch, and further split prototype error into within-episode sampling, governed by effective support, and episode-to-future shift. A one-dimensional construction admits opposite oracle rankings under the same observed information even when the positive distribution is known, establishing the need for an explicit reference assumption. On 28 later CHB–MIT seizures from 13 patients, changing only the reference moves patient-macro AP from 0.2656 to 0.3191. A source-selected linear SVM reaches 0.3175 AP, with paired AP and AUROC difference intervals against CP-0 spanning zero. These results motivate stating and justifying the reference as part of positive-only personalization.

Primary area:
learning on time series and dynamical systems
Keywords:
positive-only personalization, seizure forecasting, EEG, positive prototype, pairwise ranking, average precision, CHB-MIT, ChronoProto

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_zso9OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper is set in the context of seizure forecasting, which is basically modelled as a binary classification between preictal (roughly the 30 min period prior to a seizure) and interictal EEG segments. In this context, the paper studies the adaptation of pre-trained models to a new patient using only preictal samples preceding an initial recorded seizure, but no background (interictal) samples. The ChronoProto method proposed in the paper basically starts from a prototypical network approach pre-trained on source data. To then adapt it to the target data using only positive (preictal) samples, the prototype for the positive class is recomputed from the target samples using the frozen encoder trained on source data. The results show no significant improvement in performance compared to source-only baselines. The authors further conclude that the reference (i.e. the prototype of the negative/interictal class) affects performance.

Critical strengths

None

Critical weaknesses
  • The paper provides neither a methodological contribution nor any practicable application-specific value. The proposed ChronoProto method uses an existing approach (prototypical networks) and demonstrates no improvement over source-only baselines. The problem setting of assuming access to only preictal but no interictal data appears artificial and without any practical relevance.
  • The conclusion that the reference (prototype for the negative/interictal segments) matters is hardly surprising. The prototypes are the key constituents of prototypical networks, the distance to which directly determines the classification.
  • The paper is hard to follow, not well motivated, and lacks methodological details (see questions for details).
Questions for authors
  • Can you motivate the studied problem? Why should we care about calibration using only preictal but not interictal data? Isn't it extremely unlikely that one would have access to only preictal data in practice?
  • Can you clearly delineate in the methods section (ChronoProto) where you are simply following prior work (e.g. prototypical networks), and what (if anything) is new? Prototypical networks are acknowledged in related works, but the methods section could be improved by making these distinctions clear.

Lacking methodological details. For instance, section 3.2 introduces the source-defined representation space.

  • L169: what is meant by "positive locations"? Positive (i.e. preictal) samples?
  • L171/172: "m_s is the balanced source mean": balanced by what?
  • L173: "D_s is a diagonal matrix of scale factors computed from source data": computed how and what for?
  • L181: "balanced variance": balanced by what?

Further suggestions:

  • The word "release" is used several times in a confusing way: e.g. L015 (a protocol that releases an episode), L153 (chronological release of episode).
  • I had to read large parts of the paper just to understand the problem setting (as I summarised it above). A few introductory sentences to explain how seizure prediction is framed as binary prediction between interictal and preictal segments, what you consider the positive and negative class, etc. would improve readability. Terminology from related few shot literature (e.g. episode, support) could also be briefly introduced before it is used.
Reviewer_axU6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper first defines a task where a model can learn about a new patient using only EEG recorded before that patient’s first seizure. They then build a simple method that uses these recordings as a template to predict later seizures. The paper examines the limits of this approach: later seizures may look different, and the model has not seen what this patient’s EEG looks like outside the periods before seizures. Finally, they test the method and compare different baseline choices.

Critical strengths

The limitation of the proposed method is reported transparently.

Critical weaknesses
  1. The paper does not prove the "personalization" claim made by the paper: The paper claims that ChronoProto “personalizes a forecaster” (L148), and that the first episode “locates the patient’s positives” (L445). It also states that source-only selection is “necessary context for any personalization claim” (L358). However, ChronoProto (CP-0) performs similarly on average to Source-SVM, which uses no target-patient data (L350–354). Furthermore, the paper identifies possible reasons why personalization may fail, but does not deep dive which reasons explain the results. It therefore remains unclear when the first episode helps or when one reference should be preferred over another.

  2. Without comparison of more recent, stronger EEG foundational model baselines: There are many EEG foundation models, such as CBraMod or NeuroLM (as described in the related works). The authors should include comparisons with those stronger baselines to demonstrate the effectiveness of their findings.

  3. The reason for excluding available patient data is not convincing: The authors' defined protocol does not use the patient’s background and unlabeled EEG, even though the authors acknowledge that these recordings are often available (L38–41). This definition lets them study what one episode provides, but the paper does not show how the findings inform decisions about when to personalize a model or when additional patient data are needed.

  4. The cohort is very small: The main experiment includes 13 patients and 28 later seizures; the supplementary groups contain only two and three patients (L304–311, 417–420). Larger independent evaluations are needed to determine whether the papers' claim generalizes, for example, tuh_eeg and SeizeIT2.

  5. In a real clinical setting, the model needs to scan the EEG data to forecast the seizure; thus, in real clinical use, the model must score continuous EEG without knowing when a seizure will occur, as forecasting is not a retrospective study. The author could use the protocol presented in the paper to train the model, but using the same protocol to evaluate the model's performance cannot reflect the actual usefulness of the model, i.e., what is the false positive rate of the model when scanning long EEG.

Questions for authors
  1. What does the patient’s first episode actually add?
  2. How does ChronoProto compare with EEG foundation models used without patient-specific adaptation or with limited fine-tuning? Can these models match or outperform ChronoProto while using the same amount of patient data or less?
  3. What useful insight comes from excluding the patient’s available background and unlabeled EEG?
  4. Can ChronoProto generalize in a larger EEG cohort?
  5. What is the false positive rate of the model in scanning long EEG.
Reviewer_vDPmOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

In this paper, the authors propose ChronoProto to study patient-specific seizure forecasting. It uses a frozen source encoder and compares a first-episode prototype with a fixed reference. The authors split the pairwise ranking error into prototype error and reference mismatch, and further split prototype error into within episode sampling and episode-to-future shift. The proposed method is validated on CHB–MIT dataset including 13 subjects with 28 later seizures.

Critical strengths
  1. This paper aims to address personalized seizure prediction with limited target-domain data. The analysis may provide additional insights into the practical seizure prediction.
  2. It provides various analyses on the results interpretation.
Critical weaknesses
  1. While the research topic is of interest, the methodological novelty is limited. The advantages over existing prototype-based methods require further justification.
  2. The assumption that using only the first preictal episode and excluding all target interictal and unlabeled EEG is not sufficiently established.
  3. The baseline comparisons are insufficient. It should include more recent and SOTA methods.
  4. The performance gains over some methods are marginal.
  5. The presentation requires substantial improvement. Some expressions are wordy or uncommon in this field. Some concepts should be clearly defined when first introduced.
  6. The figures require improvement. In particular, Figure 1 contains many small labels and mathematical annotations, making the main workflow difficult to follow. The main novelty and contribution are also hard to identify in this framework figure.
Questions for authors

see above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.