acceptodds
Withdrawn from ICLR 2027

SPICE: A Subject-Preserving Counterfactual Augmentation Framework for EEG Disease Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

EEG-based disease detection often generalizes poorly to unseen subjects. A key challenge is that each subject contributes recordings with a single diagnostic label, enabling models to exploit subject identity as a shortcut for disease prediction. We propose SPICE (Subject-Preserving Inter-Class Expansion), a counterfactual augmentation framework that challenges this shortcut by generating opposite-class EEG samples while preserving subject-specific characteristics. By providing training subjects with examples of both diagnostic classes, SPICE aims to encourage learning of disease-related features that transfer across individuals. We formulate counterfactual generation as unpaired translation between healthy and disease domains and instantiate the framework using CycleGAN with subject-preservation objectives. These objectives also guide a task-specific selection policy that evaluates subject preservation and diagnostic class change while near-copies are screened out of the training view. Experiments across four EEG datasets demonstrate the potential of subject-preserving inter-class expansion to improve cross-subject disease detection, with benefits that depend on the dataset and classifier.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG, disease detection, generalization, synthetic data

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_nDB5OpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper tackles the Single-Class-Per-Subject (SCPS) problem in EEG classification, where models learn subject-specific identity shortcuts rather than generalizable disease features. The authors propose SPICE, a data augmentation framework that uses a constrained CycleGAN to generate counterfactual EEG.

Critical strengths

he paper correctly identifies the SCPS shortcut as a critical bottleneck in EEG decoding. The idea of intervening at the data level by synthesizing opposite-class signals for the same subject is a conceptually sound strategy to break this shortcut.

Critical weaknesses
  1. The paper is essentially a standard CycleGAN with identity-class as adversarial loss, which widely explored in the GAN literature.
  2. The results are mixed: out of 23 settings, only 6 show a clear improvement. The conclusion that the framework is broadly effective across cohorts is therefore overstated.
  3. Lack of ablation study for the components in the training loess.
Questions for authors
  1. Could you elaborate on the specific technical novelty of your generative framework beyond applying this standard GAN recipe to the EEG domain?
  2. Could you justify the conclusion that the framework is broadly effective across cohorts?
  3. Can you provide an ablation study removing each component in you training loess (subject loss, subject adversarial loss, spectral matching, displacement penalty) to show that they all actually drive the performance gains?
Reviewer_D5DLOpenReview ↗
Rating
1(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SPICE, a subject-preserving counterfactual augmentation framework for cross-subject EEG disease detection. It addresses the subject-identity shortcut problem in single-class-per-subject (SCPS) datasets, where each subject contributes recordings from only one diagnostic class. SPICE uses a CycleGAN-based framework with subject-preservation constraints to generate opposite-class EEG samples while retaining subject-specific characteristics. A quality-based selection mechanism further filters generated samples before incorporating them into classifier training. Experiments on four EEG datasets and multiple classification architectures under subject-disjoint evaluation show that the proposed augmentation can improve cross-subject classification performance in certain settings compared with both real-only training and matched real-data augmentation.

Critical strengths
  1. The paper addresses the subject-identity shortcut problem in EEG disease detection, which is important for improving cross-subject generalization.
  2. The proposed framework combines inter-class EEG generation with subject-preservation constraints and sample filtering, providing a reasonable way to augment training data.
Critical weaknesses
  1. The proposed method mainly combines CycleGAN with subject-preservation constraints and sample filtering. Similar ideas have been explored in previous EEG generation studies, but the paper does not clearly explain the differences.
  2. It is unclear whether the generated EEG samples can really preserve subject information while changing disease-related features. Since each subject has only one disease label, the subject encoder may also learn disease-related information.
  3. The paper does not provide sufficient ablation studies to show the contribution of each component. Direct comparisons with standard CycleGAN and other EEG augmentation methods are also missing.
  4. The paper does not directly show that SPICE reduces the model's reliance on subject identity. The improvements are also inconsistent across datasets and models, and some gains may be related to training stability rather than better generalization.
Questions for authors
  1. How does SPICE differ from existing counterfactual EEG generation and subject-conditioned augmentation methods? What is the main technical novelty beyond combining existing components?
  2. How do the authors verify that the generated EEG samples preserve subject-specific information while changing disease-related features? Given that subject identity and disease labels are correlated in SCPS datasets, could the identity encoder also capture disease-related information?
  3. Can the authors provide ablation results for the main components and comparisons with standard CycleGAN or other EEG augmentation methods? How much does each component contribute to the final performance?
  4. Is there any direct evidence that SPICE reduces the classifier's reliance on subject identity? Could some of the reported improvements be explained by better training stability rather than reduced identity shortcuts?
Reviewer_3amqOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SPICE proposes an EEG augmentation framework that generates examples with the opposite diagnostic label while seeking to preserve characteristics of the source subject. It combines unpaired translation, subject consistency constraints, and candidate filtering to introduce diagnostic variation within subjects. Experiments evaluate whether these examples improve disease classification on unseen subjects beyond adding repeated real recordings.

Critical strengths
  1. The paper addresses an important problem in clinical EEG: models may rely on individual characteristics that do not generalize to unseen subjects.
  2. The evaluation covers several clinical cohorts and classifiers, separates subjects across training and evaluation, and includes a control for the number and class composition of added examples.
Critical weaknesses
  1. The experiments compare SPICE with real data and matched resampling, but include no alternative EEG augmentation or generative augmentation methods. These comparisons assess gains over repeated real samples but do not establish the practical value of SPICE relative to existing augmentation approaches.
  2. The central goal of preserving personal characteristics while changing diagnostic characteristics is not independently validated. The same subject encoder guides generation and candidate selection and may retain diagnostic information, while diagnostic classifiers provide acceptance criteria after generation. These proxies do not establish that the intended preservation and change occur together, leaving the counterfactual interpretation insufficiently supported.
  3. The experiments do not isolate whether the subject preservation design accounts for the downstream benefit. Section 4.4 does not ablate the preservation objectives or associated filtering, and there is no direct assessment of reduced reliance on subject identity. The observed gains therefore cannot be attributed specifically to the proposed mechanism of introducing diagnostic variation within subjects.
  4. The results vary across cohorts and classifiers, but the experiments do not test the conditions under which the proposed mechanism should help. Controlled variation in subject shortcuts or in the relationship between personal and diagnostic characteristics would clarify the method’s scope and failure conditions.
Questions for authors
  1. How does SPICE compare with representative EEG augmentations and a comparable generator without the subject preservation objectives and filtering?
  2. What independent evidence shows that generated signals preserve personal characteristics while changing the intended diagnostic characteristics, beyond satisfying the criteria used during generation and selection?
  3. Under controlled changes in subject shortcuts, does SPICE reduce the classifier’s reliance on those shortcuts, and does this reduction explain its generalization benefit?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Thank the revierwers for their comments

We thank the reviewers for their thoughtful feedback. We agree that the current results do not demonstrate consistent gains across datasets. We have therefore decided to withdraw the submission. The reviewers’ comments on the lack of ablation studies and the limited evidence that our method preserves subject information will guide our future work.

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

Related papers

Discussion (0)

Sign in to comment.