acceptodds
Under review as a conference paper at ICLR 2027

PASF: Persistence-Aware Spectral Filtering for Test-Time Adaptation in Sleep Staging

Abstract

Physiological signal models generalize poorly across devices, sensor configurations, and recording protocols. Such acquisition shifts frequently manifest in the frequency domain, motivating spectral adaptation at test time. However, a recording's marginal spectrum reflects not only acquisition conditions but also the task-relevant physiological states it contains and their prevalence. Wholesale alignment can therefore suppress state-dependent physiology alongside domain nuisance, particularly under label shift. We study source-free test-time adaptation and identify cross-state-persistent spectral shifts as a conservative correction target. These persistent shifts remain consistent across physiological states, while state-specific disagreement serves as counterevidence against full correction. We instantiate this principle as Persistence-Aware Spectral Filtering (PASF), which estimates probability-weighted state-conditioned spectral residuals relative to compact source templates. Their shared component shapes a recording-specific, phase-preserving filter, while the shared-energy fraction gates its strength. PASF processes each complete recording in two passes without target labels, raw source recordings, model updates, or gradients. Across four sleep-staging datasets comprising 740 recordings, PASF achieves four-domain mean balanced accuracies of 75.56% with U-Sleep and 74.30% with a CNN–Transformer. These results improve on source-only inference by 3.79 and 5.78 percentage points, respectively. Controlled comparisons show reduced composition-induced false correction and improved recovery under shared spectral distortion. Analyses of real recordings support cross-state persistence, while experiments beyond sleep staging identify the applicability and limits of the shared-shift principle.

Primary area:
applications to neuroscience & cognitive science
Keywords:
test-time adaptation, source-free adaptation, spectral filtering, distribution shift, sleep staging, electroencephalography

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_TAyNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EEG recordings vary across datasets, and differences in sleep-stage proportions can affect spectral alignment during adaptation. This paper proposes PASF, which uses predicted sleep-stage probabilities and source-stage spectral templates to estimate a recording-specific filter at test time. The method corrects spectral differences shared across stages while keeping the prediction model fixed. The paper evaluates this approach across four sleep datasets and two model architectures, with additional experiments examining its mechanism and limitations.

Critical strengths

The cross-dataset evaluations and controlled experiments provide useful evidence for the proposed approach. The method improves average performance without updating model parameters or requiring target labels. The detailed ablations and reporting of negative results also help clarify when spectral correction may be useful and where its assumptions fail.

Critical weaknesses
  1. The practical contribution of target-stage conditioning is not clearly established by the main results. How much the of the real-data improvements comes from target-stage conditioning?
  2. The assumption that differences shared across stages are suitable for correction needs clearer conditions of applicability.
  3. The evaluation does not establish whether the gains extend to stronger pretrained sleep models which targets the generalization problem.
Questions for authors
  1. Why frequency-domain is the best part to learn the adaptation, is there any experiments to support the claim or just follow the previous research?
  2. How these task-relevant physiological state will affect the adaptation?
  3. In line 81, how to identify the task-relevant physiological states from the training data using state labels?
  4. In the related work, i understand the overcorrection under label shift? how about other real-world shift problems and how this framework can solve other shift conditions?
  5. There are a lot of sleep foundation models such as sleepFM[1], OSF[2], why not choosing them as the evaluation models or such generalization problem is less challenging for those settings?

[1]. Thapa, Rahul et al. “A multimodal sleep foundation model for disease prediction.” Nature Medicine 32 (2026): 752 - 762. [2]. Shuai, Zi-Tao et al. “OSF: On Pre-training and Scaling of Sleep Foundation Models.” ICML 2026

1 comment
AuthorsOpenReview ↗

Response to the Reviewer TAyN

Response to the Reviewer

We thank the reviewer for the constructive comments. We appreciate the opportunity to clarify the choice of the frequency domain, the role of physiological states, the scope of the correction framework, and its applicability to sleep foundation models.

  • The role of target-stage conditioning. Our core contribution is to identify and selectively correct recording-level spectral shifts shared across states. Uniform retains key elements of this framework and also targets shifts that persist across states; its strong performance therefore supports the broader correction approach. However, our controlled mechanistic experiments show that this variant remains sensitive to the target-state distribution, motivating target-stage conditioning to improve robustness to changes in sleep architecture.
  • Conditions for correcting shared shifts. Since our objective is physiological-state classification, our rationale is that a spectral shift affecting all states equally does not directly provide cues for distinguishing those states. Such a shift can nevertheless interfere with the transfer of the source classifier across domains. We therefore consider selective correction appropriate and warranted for this state-classification task. We acknowledge that the same shared shift may carry useful information for other tasks.
  • Evaluation with sleep foundation models. We have additionally evaluated PASF with SleepFM and OSF on four target datasets. PASF improves mean balanced accuracy by 2.60 and 2.29 percentage points, respectively, supporting its additional value on top of the evaluated foundation models.

We address the specific questions below.

Q1. Why perform adaptation in the frequency domain?

Response: Our choice of the frequency domain is supported by both theoretical motivation and prior empirical evidence. CMMN connects spectral alignment to convolutional corrections under stationary-signal assumptions and demonstrates improvements over conventional normalization in sleep staging [1]. Moreover, FMM directly compares Fourier-space and input-space moment matching using the same architecture: on Sleep-EDF, Fourier-space matching achieves 74.1% accuracy versus 65.4% for input-space matching [2]. Building on this evidence, PASF focuses on identifying which spectral discrepancies should be corrected.

[1] Gnassounou, T., Flamary, R., and Gramfort, A. Convolutional Monge Mapping Normalization for Learning on Sleep Data. In Advances in Neural Information Processing Systems, vol. 36, 2023.

[2] Krishnamachari, K., Ng, S.-K., and Foo, C.-S. Mitigating Real-World Distribution Shifts in the Fourier Domain. Transactions on Machine Learning Research, 2023.

Q2. How these task-relevant physiological state will affect the adaptation?

Response: Different sleep stages have distinct spectral characteristics, so changes in stage proportions can alter the overall spectrum even when acquisition conditions remain unchanged. Ignoring these states may lead to physiological variation being mistaken for domain-related distortion. PASF uses predicted stage probabilities to compare target signals with stage-specific source templates, then identifies the component shared across the resulting state-conditioned residuals. This helps distinguish spectral differences due to stage-composition changes from candidate recording-level shifts.

Q3. How to identify the task-relevant physiological states from the training data using state labels?

Response: We use the provided source labels to group EEG epochs into the five sleep stages and compute the mean log-power spectrum within each group as its initial state-specific template.

Q4. Which real-world shifts can PASF address beyond label shift?

Response: PASF targets the structure of a shift rather than its origin. Whether a shift arises from devices, populations, or environmental conditions, PASF aims to correct its spectral component that remains stable within a recording and is shared across states. Its scope therefore extends beyond label shift, but does not cover shifts outside this structural assumption.

Q5. Does PASF provide additional benefits for sleep foundation models?

Response: We evaluated PASF with SleepFM and OSF on four target datasets. PASF improves mean BAcc by 2.60 and 2.29 percentage points, respectively. Each model improves on three of the four datasets, with ISRUC as the exception. The full per-dataset results are in item 4, “Stronger pretrained backbones,” of our Response to the Area Chair (1/2).

Reviewer_QgbyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose PASF, a test-time adaptation method for sleep staging that keeps the source classifier frozen. PASF estimates stage-conditioned spectral residuals and extracts their shared component. A shared-energy coefficient controls the strength of a recording-specific, phase-preserving filter. Across four datasets and two backbones, PASF improves average balanced accuracy over source-only inference and several adaptation baselines. Additional experiments examine stage-composition shifts, probability quality, negative transfer, and applicability beyond sleep staging.

Critical strengths
  • The paper identifies an important limitation of marginal spectral alignment, that differences in sleep-stage composition can be mistaken for acquisition shifts. PASF provides a computationally simple correction procedure without test-time gradient updates.
  • Experiments include diverse settings, backbones, datasets, and uncertainty estimates. Controlled interventions and detailed ablations help investigate the proposed mechanism.
  • The paper explicitly reports negative transfer, domain-dependent benefits of stage conditioning, and unsuccessful transfer settings rather than presenting adaptation as uniformly beneficial.
Critical weaknesses
  • The Uniform variant removes target-stage information yet retains most of the improvement, outperforming PASF with U-Sleep and trailing it only slightly with the CNN–Transformer. This weakens the evidence that estimating target cross-state persistence is responsible for the main gains.
  • Spectral differences shared across stages can also reflect meaningful biological variation. The decomposition is mathematically valid, but its interpretation as a correction target remains an assumption. Synthetic distortion recovery and improved staging accuracy do not establish preservation of clinically relevant information.
  • The method uses probabilities from the same source classifier affected by domain shift. Systematic stage confusion can contaminate the estimated residuals, and neither robust aggregation nor the shared-energy gate guarantees protection against such errors.
Questions for authors
  1. The Uniform variant retains most of PASF’s gains and outperforms it with U-Sleep. Beyond the controlled composition-shift experiments, when does target-stage conditioning provide a meaningful advantage over this simpler variant?
  2. While in current era, many robustness or transferability problems are generally solved by scaling model size and data size. For the aquisition shift problem, does PASF improve upon a strong PSG foundation models [1][2] that are trained on more diverse data and are larger in size? Would be good if the authors can test, or discuss, if not enough resources, the contribution of this work under this context.

[1] Xu, Zongzhe, et al. "Sleeplm: Natural-language intelligence for human sleep." arXiv preprint arXiv:2602.23605 (2026). [2] Thapa, Rahul, et al. "SleepFM: multi-modal representation learning for sleep across brain activity, ECG and respiratory signals." arXiv preprint arXiv:2405.17766 (2024).

1 comment
AuthorsOpenReview ↗

Response to the Reviewer Qgby

Response to the Reviewer Qgby

We thank the reviewer for recognizing the practical motivation of our work, the breadth of our evaluation, and our explicit reporting of negative transfer and unsuccessful transfer settings. We would like to clarify three aspects of our contribution and its limitations.

  • The role of target-stage conditioning. Our core contribution is to identify and selectively correct recording-level spectral shifts shared across states. Uniform retains key elements of this framework and also targets shifts that persist across states; its strong performance therefore supports the key idea. However, our controlled mechanistic experiments show that this variant remains sensitive to the target-state distribution, motivating the introduction of target-stage conditioning to improve robustness to changes in sleep architecture.
  • The interpretation of shared spectral shifts. We acknowledge that shared spectral shifts may partly reflect physiological variation. For sleep staging, our rationale is that, insofar as a shift affects all stages equally, it does not directly provide information for distinguishing those stages. But such a shift can nevertheless affect the transfer of the source model across domains, motivating its selective correction. We do not rule out that these physiological shifts may carry clinically useful information for purposes beyond sleep staging.
  • Dependence on source-model probabilities. The Uniform ablation shows that most of the performance gains can be retained without using the source model’s target-stage probabilities to estimate the correction. This supports the view that the main benefit of identifying and correcting recording-level shared spectral shifts does not depend solely on accurate target-stage predictions. Target-stage conditioning is introduced to improve robustness to stage-composition changes.

We address the reviewer’s specific questions below, including additional results with sleep foundation models.

Q1. The advantage of target-stage conditioning

The Uniform variant retains most of PASF’s gains and outperforms it with U-Sleep. Beyond the controlled composition-shift experiments, when does target-stage conditioning provide a meaningful advantage over this simpler variant?

Response: Uniform is an ablation within our framework that retains source-stage templates, an equal-stage reference, and adaptive correction. Its strong performance supports our central approach—shifting the correction target from wholesale marginal spectral alignment to selective correction of recording-level shared spectral shifts. The additional value of target-stage conditioning is its robustness to changes in sleep-stage composition. This is particularly relevant when deploying across populations with systematically different stage proportions. With acquisition distortion and evaluation recordings fixed, increasing N3 in the adaptation context reduces Uniform’s balanced-accuracy gains by 1.57 and 1.34 percentage points for U-Sleep and CNN–Transformer, respectively, while PASF’s gains remain nearly unchanged. These results motivate stage conditioning as a way to preserve classification gains under changes in sleep architecture.

Q2. Complementarity with large-scale PSG foundation models

While in current era, many robustness or transferability problems are generally solved by scaling model size and data size. For the aquisition shift problem, does PASF improve upon a strong PSG foundation models [1][2] that are trained on more diverse data and are larger in size? Would be good if the authors can test, or discuss, if not enough resources, the contribution of this work under this context.

Response: Thank you for this suggestion. We evaluated PASF with SleepFM and OSF on four target datasets, as shown below. PASF improves mean BACC by 2.60 and 2.29 percentage points, respectively, with gains on three of four datasets for each model. These results demonstrate that PASF provides additional benefits on top of the evaluated foundation models, supporting its complementarity with large-scale pretraining.

ModelTarget datasetBaseline BACC (%)+PASF BACC (%)Δ (pp)
SleepFMHMC67.8569.84+1.99
SleepFMISRUC64.9264.38−0.54
SleepFMSHHS69.2169.64+0.43
SleepFMSleep-EDF52.3160.83+8.52
SleepFMMean63.5766.17+2.60
OSFHMC73.6776.40+2.73
OSFISRUC71.3470.08−1.26
OSFSHHS72.7473.72+0.98
OSFSleep-EDF61.2767.97+6.70
OSFMean69.7672.04+2.29
Reviewer_AwVaOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PASF, a test-time adaptation method for physiological signals, mainly evaluated on cross-domain sleep staging. The key idea is to estimate state-conditioned spectral residuals and identify the component shared across physiological states as a potentially correctable recording-level shift. The shared component is then used to construct a phase-preserving spectral filter, with its strength controlled by a persistence score. Experiments cover four sleep datasets, two backbones, several TTA baselines, extensive ablations, controlled perturbations, and additional EEG tasks. The method shows strong average improvements over source-only inference and several existing adaptation methods.

Critical strengths

The motivation is clear and practically relevant. The paper correctly points out that marginal spectral alignment may remove not only acquisition-related variation but also task-relevant physiological differences, especially when the target state distribution changes.

The experimental evaluation is unusually thorough. The paper evaluates four sleep-staging datasets, two different backbones, multiple spectral and model-level adaptation baselines, controlled interventions, and several additional EEG settings.

The average empirical gains are strong. PASF improves four-domain balanced accuracy over source-only by 3.79 and 5.78 percentage points for the two tested backbones, respectively.

Critical weaknesses

Limited contribution of the core mechanism. The Uniform-responsibility ablation removes target-state conditioning but retains nearly all the performance gains, even outperforming PASF on U-Sleep. This raises concerns about whether cross-state persistence is truly responsible for the improvements.

Limited methodological novelty and theoretical justification. The method mainly combines state-conditioned residual averaging with an energy-based correction gate. Moreover, cross-state persistence does not necessarily indicate acquisition-related nuisance, leaving the underlying assumption insufficiently justified.

Limited reliability and generalizability. Around 40% of recordings experience negative transfer. The method also requires complete recordings for offline adaptation and shows inconsistent improvements on other EEG tasks, limiting its practical applicability.

Questions for authors

What advantage does target-state conditioning provide over the simpler Uniform-responsibility variant, which retains nearly all the gains?

What justifies using cross-state persistence to guide correction when persistent shifts may also reflect meaningful physiological variation?

Can a label-free criterion identify recordings likely to suffer negative transfer and determine when PASF should abstain from adaptation?

1 comment
AuthorsOpenReview ↗

Response to the Reviewer AwVa

Response to the Reviewer AwVa

We thank the reviewer for recognizing the practical relevance of our motivation, the comprehensiveness of our evaluation, and the substantial average improvements. We would like to clarify the scope of our contribution and how the identified limitations should be interpreted.

  • Core contribution and the role of target-state conditioning. The central idea is to move from direct spectral alignment to identifying and selectively correcting spectral shifts that persist across states. The effectiveness of the Uniform-responsibility variant supports this idea. However, this variant remains sensitive to the target-state distribution, motivating the introduction of target-state conditioning to improve robustness to changes in sleep architecture.
  • Justification for correcting shared shifts. We acknowledge that shifts shared across states may reflect factors beyond acquisition noise. Our assumption is that, insofar as these shifts affect all states equally, they do not directly provide information for distinguishing those states. Nevertheless, they can affect the transfer of a trained model across domains, which motivates their selective correction.
  • Reliability and broader applicability. Because PASF operates at the input level, its ability to control downstream negative transfer is limited. In the held-out U-Sleep evaluation (seed 42), the source-selected fallback rule retained about 92% of the BAcc gain from applying PASF to all recordings and reduced the share of all recordings with BAcc losses greater than one point from 22.46% to 9.50% (see Q3). The improvements observed on tasks beyond sleep staging also suggest potential applicability beyond our primary setting. For these tasks, exploring representations in which shared shifts can be more effectively identified and corrected is a promising direction for extending this approach.

We address the reviewer’s specific questions below.

Q1. The advantage of target-state conditioning

What advantage does target-state conditioning provide over the simpler Uniform-responsibility variant, which retains nearly all the gains?

Response: Uniform is an ablation within our framework that retains source-stage templates, an equal-stage reference, and adaptive correction. Its strong performance supports our central approach—shifting the correction target from wholesale marginal spectral alignment to selective correction of recording-level shared spectral shifts. Target-stage conditioning further improves robustness to stage-composition changes. With acquisition distortion and evaluation recordings fixed, increasing N3 in the adaptation context reduces Uniform’s balanced-accuracy gains by 1.57 and 1.34 percentage points for U-Sleep and CNN–Transformer, respectively, while PASF’s gains remain nearly unchanged. Thus, when stage proportions differ systematically across deployment populations, target-stage conditioning helps preserve classification gains under changes in sleep architecture, particularly shifts in sleep-stage proportions across deployment populations.

Q2. Justification for cross-state persistence as a correction criterion

What justifies using cross-state persistence to guide correction when persistent shifts may also reflect meaningful physiological variation?

Response: We agree that cross-state shared shifts may include physiological variation. For sleep staging, our focus is their contribution to distinguishing stages. Under the shared additive log-spectral shift model, a component that affects all stages equally does not directly contribute to their relative spectral differences. This motivates correcting the shared component while preserving stage-specific discriminative information. Cross-state persistence therefore provides a task-relevant correction criterion without requiring every shared shift to be attributed to acquisition noise. Our controlled experiments and cross-domain classification results support this criterion under the tested conditions.

Q3. Label-free assessment of negative transfer and abstention

Can a label-free criterion identify recordings likely to suffer negative transfer and determine when PASF should abstain from adaptation?

Response: We tested a label-free rule that applies PASF when γ≥0.8\gamma\ge0.8 and uses Source-only predictions otherwise. With the threshold selected on source validation, this rule retained about 92% of the BAcc gain of Always PASF on the held-out U-Sleep evaluation (seed 42). Across all recordings, BAcc negative transfer fell from 35.13% to 13.24%, and losses exceeding 1 point fell from 22.46% to 9.50%. See item 3, “Negative transfer and scope,” in our Response to the Area Chair (1/2) for the full comparison.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.