acceptodds
Under review as a conference paper at ICLR 2027

Disentangling OOD Detection and Model Uncertainty in High-Risk EEG Applications

Abstract

Machine learning models for electroencephalography (EEG) analysis show great promise across a wide range of applications, but their deployment in high-risk domains is hindered by their vulnerability to distribution shifts. Encountering out-of-distribution (OOD) data can lead to catastrophic, overconfident predictive failures. While OOD detection methods can mitigate these risks, they remain heavily under-explored for EEG. Moreover, evaluations in the broader literature typically assess OOD detection performance in isolation, ignoring their practical impact on downstream applications. To bridge this gap, we introduce a benchmark for EEG OOD detection, evaluate a broad range of methods, and investigate their value in two downstream clinical prediction task. Our results disentangle OOD detection and model uncertainty estimation capabilities, which are frequently conflated in the literature, provide actionable insights about the current state of the art for EEG OOD detection and model uncertainty estimation, and demonstrate how complementary methods for both aspects can be combined to form a robust safety net for the deployment of machine learning models in real-world applications.

Primary area:
applications to neuroscience & cognitive science
Keywords:
OOD Detection, Model Uncertainty, AI Safety, Robustness, Reliability, Electroencephalography, EEG

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_bb39OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a benchmark for EEG OOD detection, evaluate a broad range of methods, and investigate their value in two downstream clinical prediction tasks. In its framework, the paper uses SITN for OOD detection and calibrated MSP as a proxy for predictive uncertainty.

Critical strengths
  1. The analysis is interesting. The comparison between mixed ID/OOD data and ID-only data in Figure 3 reveals an interesting difference: MSP can be informative about classification accuracy despite poorly separating ID from OOD, whereas SITN effectively detects the evaluated distribution shift but provides less information about accuracy within most ID samples. This offers a useful empirical perspective on the different roles of these scores.

  2. Thorough analysis in TUAB dataset. The paper provides extensive supplementary analyses, including hyperparameter sensitivity, comparisons across classifier architectures, sample-size effects, and detailed downstream results across perturbation conditions.

Critical weaknesses
  1. Limited methodological novelty and an overstated title. The study primarily evaluates existing methods, and the proposed method has limited methodological innovation. The title’s use of “disentangling” appears stronger than the experimental evidence supports. The main experiments show that SITN and MSP differ in their ability to detect perturbed inputs and predict classification errors. However, model uncertainty is primarily operationalized through MSP-based confidence, without a clear definition of the notion of uncertainty being assessed. The results support a distinction between OOD detection and confidence-based error prediction, but the broader claim of disentangling OOD detection and model uncertainty estimation requires clearer conceptual definitions and more targeted empirical validation.
  2. Insufficient experimental breadth for a benchmark contribution. The coverage of datasets remains limited relative to the paper’s broad claims. The evaluation does not include representative EEG foundation models, such as LaBraM and CBraMod, leaving it unclear whether the main conclusions extend to representations learned through large-scale pretraining. Furthermore, OOD detection performance is assessed primarily using AUROC. Reporting complementary metrics, such as AUPRC, balanced accuracy would help. Finally, results across multiple independent training seeds are not reported, making it difficult to assess the stability of the findings.
Questions for authors

Could the authors clarify the justification for treating the proposed perturbations, particularly at low severity, as OOD conditions? Some perturbations may produce variations already present in ID EEG, so applying a transformation does not automatically establish a meaningful distribution shift.

1 comment

Thank you for your thorough review. We appreciate that you found our analysis interesting and useful.

Notion of uncertainty

While we would argue that the results and their discussion clearly outline the notion of uncertainty that we observe to be reflected in scores from discriminative methods like the MSP, we agree that “model uncertainty” may not be the right terminology to use here.

We have updated the terminology to “predictive uncertainty” in accordance with the terminology of Hüllermeier and Waegeman [1] along with the following description to clarify it:

Discriminative methods like MSP, on the other hand, are not suitable for OOD detection, but reflect predictive uncertainty. Even with ample exposure to similar samples during training, certain cases can be ambiguous and difficult for the model to classify. MSP scores can capture such cases and provide valuable information about how well the model is likely to perform across different in-distribution samples.

With this clarification in terminiology, we believe our experimental evidence appropriately supports the paper's claims.

[1] Hüllermeier, E., Waegeman, W. Aleatoric and epistemic uncertainty in machine learning: an introduction to concepts and methods. Mach Learn 110, 457–506 (2021). https://doi.org/10.1007/s10994-021-05946-3

Further evaluations

We have added extended results tables with additional metrics (AUPR and FPR@95TPR) in Appendices A and B.

Regarding training seed variability, methods are compared using the same underlying models. This is a different situation than comparing separately trained models where training variability can obscure performance comparisons. Moreover, results are consistent across the separate training runs for the different datasets as well as multiple training runs on different data subsets (Appendix E).

Unfortunately, foundation models like LaBraM and REVE were trained with substantially different data preprocessing setups than what was used in our experiments (different sampling rate, different filtering). Clean comparisons to these models would therefore either require re-training of the foundation models with matched preprocessing or considerable adjustments to the experimental setup (re-definition of perturbation severities for resampling and filtering perturbations; re-evaluation of the current models with matched preprocessing). While we agree that evaluations with foundation models could be interesting, they are not required to support the claims of our work and could be better explored in future work.

Justification for perturbations at low severity levels

We agree that in principle certain small perturbations could simulate some natural variation. However, this is unlikely to be the case for the specific perturbations and lowest severity levels chosen in our experiments. For instance, even a single channel swap would not be consistent with natural variation. Importantly, our results are robust in that even if the lowest severity level was ignored, the conclusions are unchanged.

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

The work benchmarks discriminative and generative OOD scores under controlled EEG perturbations, with observations that generative methods usually distinguish perturbed from unperturbed inputs more effectively, whereas classifier confidence better predicts errors within in distribution data. The method then combines SITN and calibrated maximum softmax probability (MSP) for selective prediction.

Critical strengths

It is an important practical question, connecting OOD detection to downstream prediction quality is more informative than reporting detection AUROC alone.

The work involves substantial empirical coverage. The study uses two large clinical datasets, multiple detection methods, with five perturbation families.

Useful supplementary controls. The appendices include additional classifier architectures, hyperparameter sensitivity, training data scaling, and cross participant experiments.

Critical weaknesses

The combined rejection system is not evaluated at matched coverage. The combination has the highest accuracy, but also rejects the most inputs. This is encouraging but don't provide sufficient evidence that combining the signals produces the best accuracy coverage tradeoff.

The paper repeatedly states that discriminative methods are incapable of detecting OOD data or that only generative methods provide protection. However, its own supplementary results show meaningful exceptions: Mahalanobis distance achieves 0.925 AUROC on CAUEEG, Energy achieves 0.792 AUROC for the strongest high-pass perturbation on CAUEEG. Below-chance AUROC also needs careful interpretation: it indicates an incorrectly oriented ranking, which can still contain distributional information.

How much of the generative advantage remains against simple spectral or covariance detectors?

Questions for authors

Please see above

1 comment

Thank you for your constructive review. We are glad that you appreciate the empirical breadth and the practical relevance of our work.

Evaluation at matched coverage

Thank you for the comment. We conducted a control analysis where we evaluated the performance of the individual methods with coverage matched to the combination approach. These results confirm that the combined approach achieves the best performance in both settings (ID-only and ID+OOD). When evaluated on ID-only data, MSP alone achieved the same performance as the combined approach at matched coverage. This is expected because the OOD detection method is not adding value if there are no OOD samples to be flagged. However, on ID+OOD data, the combined approach performed better than either of the individual methods at matched coverage.

We added these results in Appendix J in the updated manuscript and print them here for convenience:

Rejection MethodID Only CoverageID Only AccuracyID and OOD CoverageID and OOD Accuracy
Combined (SITN + MSP)86.9%84.1%43.7%84.1%
SITN (Matched Coverage)86.9%80.7%43.7%80.7%
MSP (Matched Coverage)86.9%84.1%43.7%74.9%

Exceptions of good discriminative performance and below-chance AUROC results

Indeed, individual discriminative methods achieved good performance on certain perturbations. This is not only the case in the supplement, but also in Table 1, where Energy achieves an AUROC of 0.684 for the highest severity of the high-pass perturbation. However, the point is that none of the discriminative methods perform well consistently across different perturbations. Energy in the aforementioned example hovers around random performance for most of the other perturbations and severity levels. Similarly, the Mahalanobis distance works well for the re-referencing perturbation and higher severity levels of the channel shuffle perturbation, but it performs at or below chance level for other perturbations.

We also agree that below-chance AUROC values can indicate discriminative information in the score. However, a method where the interpretation of what is OOD changes per case has little practical value. As we do not assume access to OOD training or validation data, there is no way of knowing the right direction. The directionality of the scores (whether a higher score means more ID or OOD) is also clearly specified for each method; the way they are benchmarked here aligns with their intended use.

Spectral or covariance detectors

While dedicated methods could be engineered to detect the specific shifts introduced by our perturbations based on particular power spectrum or channel covariance features, we see the main value of OOD detection methods as protecting against unforeseen shifts. Because of that and because there are, to our knowledge, no existing EEG OOD detection methods based on power spectrum or covariance features, we did not evaluate such baselines.

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

The paper introduces a OOD benchmark for EEG data. Addtiionallly, shows empirically that generative OOD methods are good at detecting distributional shift, and discriminative ones are good at finding uncertainty on hard in-distribution data. The experiment compare 8 methods on two datasets.

Critical strengths
  1. Creating a benchmark sets the tone for the research in this area. Timely contribution. 2.Paper uses the standard datasets and perturbs them with a diverse set of edits and severity controls, to introduce a variety of OOD scenarios (there’s is a flip side to it, discussed later.), which is useful to convert any dataset to study OOD problem.
Critical weaknesses
  1. I think in terms of contribution the paper doesn’t introduce any new insight. The disentanglement on generative vs discriminative is entirely empirical and that correlation may just be due to methods considered. Moreover, it is by definition that density models (generative) detect distributional mismatch, and classifier scores/logits may hard instance. Therefore, the disentanglement is just repackaging in my view.
  2. If I discount disentanglement, then the paper doesn’t have much apart from applying existing methods on two datasets, without providing some useful/practical insight.
  3. The perturbation introduced surely give OOD data. But how reflective is that of real world drift?
Questions for authors

None

2 comments

Evidence for disentanglement and new insights

"The disentanglement on generative vs discriminative is entirely empirical and that correlation may just be due to methods considered."

We show very consistent results across different discriminative methods, datasets, tasks, and model architectures. The included discriminative methods are highly influential reference methods, not niche cherry-picked approaches. We have a broad and thorough empirical setup (as acknowledged by reviewer Truz). We also discuss theoretical reasons for the difference between discriminative and generative methods.

"Moreover, it is by definition that density models (generative) detect distributional mismatch, and classifier scores/logits may hard instance. Therefore, the disentanglement is just repackaging in my view."

Importantly, the included discriminative methods are widely established as OOD detection methods. The most widely cited OOD detection benchmarks (OpenOOD) specifically benchmark and promote these methods as effective methods for out-of-distribution detection. This is what researchers and practitioners refer to when selecting OOD methods for their work. Our results provide strong contradictory evidence and show that these methods are in fact not effective for OOD detection, but reflect predictive uncertainty. Therefore, we respectfully disagree with the framing of our contributions not introducing any new insights or just “repackaging”. Our results do not agree with widely held beliefs about these methods, so they are certainly not reiterating them.

Contributions beyond disentanglement

We argue above why the disentanglement claims are warranted and relevant.

We would also like to clarify that our work provides further contributions beyond the disentanglement (listed as bullet points in the introduction section) and that these have practical value. First, we do not just evaluate existing datasets. We introduce a suite of controlled perturbations to enable controlled OOD detection evaluations for EEG in the first place. Furthermore, we evaluate OOD detection for clinical EEG data and we go beyond standard OOD detection metrics to also evaluate their effect on the actual downstream clinical prediction tasks. These evaluations are directly relevant for researchers and practitioners working with (clinical) EEG data. Finally, we demonstrate how methods for predictive uncertainty and OOD detection can be combined to safeguard against both failure cases in a system.

Relevance of the perturbations

We discuss and provide examples for the plausibility of the proposed perturbations in Section 3.1 – they can and do actually happen in practice. More broadly, the point is not to protect against these specific perturbations, but against unforeseen shifts – the methods were not designed in any way to be able to detect these specific shifts; we use them as a test bed. Finally, we also discuss the difficulty of constructing OOD detection benchmarks for EEG data, provide further exploratory analysis with cross-dataset setups and population differences, and provide advice for future work along these lines.

Thank you for the clarifications, and your thoughtful response. I think it's fair that discriminative methods are widely used in OOD baselines (OpenOOD), and therefore eliciting their failure modes has practical implications. I am 100% onboard with this. On the other hand, though deviating from becnhmarks like OpenOOD, there is difference between density based and confidence based detection, purely from first principles perspective.

Having said that, I am willing to revise my score upwards after reviewer discussions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.