acceptodds
Withdrawn from ICLR 2027

CtxSleepNet: Context-Aware Multimodal Sleep Staging With Label-Efficient Domain Adaptation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Generalization across recording channels, acquisition devices, and clinical centers remains a major challenge in multimodal sleep staging. We propose CtxSleepNet, a context-aware framework that integrates multimodal representation learning with adaptation tailored to the available target-domain label budget. A hierarchical CNN–GRU–Transformer encoder captures local morphology, short-term dynamics, and long-range dependencies, while an LSTM-based cross-epoch module incorporates neighboring context to refine individual epoch predictions. For label-free transfer, we introduce recording-wise domain statistics alignment, combining stage reweighting with batch normalization correction relative to source references. This strategy exploits unlabeled target data without additional pretraining or gradient-based parameter updates. For adaptation with limited labels, we develop typicality–certainty selection (TypCert), which combines target-subject selection based on stage-distribution typicality and predictive certainty with budget-dependent parameter updating. By restricting updates at low label budgets and enabling full-network fine-tuning at higher budgets, this strategy balances estimation variance and model flexibility. Experiments on four public datasets demonstrate that CtxSleepNet consistently outperforms competing methods in-domain, achieving accuracies of 79.6%–87.3%. Across six transfer scenarios spanning channel, device, and center differences, domain statistics alignment improves mean zero-shot accuracy from 68.2% to 74.4%, a gain of 6.2 percentage points. With 10% labeled target data, mean accuracy reaches 79.6%, only 2.1 percentage points below the in-domain reference of 81.7%. These results support the effectiveness of contextual multimodal modeling and complementary label-free and label-efficient adaptation for sleep staging across heterogeneous recording environments.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Multimodal Sleep Staging, Context-Aware, Label-Efficient, Domain Generalization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tY6POpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CtxSleepNet, a multimodal sleep staging framework designed to improve cross-domain generalization. The model combines CNN, GRU, Transformer, and LSTM to extract features from EEG and EOG signals and capture temporal dependencies across sleep epochs. The authors also introduce TypCert to select target subjects for fine-tuning under limited labeling budgets and a domain statistics alignment strategy to improve zero-shot performance.

Critical strengths
  1. The experiments are comprehensive, including comparisons with existing methods, cross-dataset evaluations, ablation studies, and sensitivity analysis.

  2. The proposed method achieves promising cross-domain performance, improving the average zero-shot accuracy from 68.2% to 74.4% with domain adaptation.

Critical weaknesses
  1. The methodological novelty is limited. The backbone mainly combines existing CNN, GRU, Transformer, and LSTM modules. Similar designs for epoch-level feature extraction and sequence modeling have already been explored in previous sleep staging studies.

  2. The model is relatively complex, but the paper lacks comparisons with baselines in terms of FLOPs, parameter counts, and inference time. Although the model achieves good performance, the additional computational cost should also be evaluated.

Questions for authors

The ablation studies could be more comprehensive. While the authors evaluate different backbone components, more detailed ablations of the domain adaptation strategy are needed. Since DSA has already been studied in previous work, the authors should clarify how much improvement comes from stage reweighting and recording-wise BN correction. Similarly, TypCert is only compared with random selection, without separately evaluating the contributions of Typicality and Certainty.

4 comments

We thank the reviewer for the constructive comments. All new experiments reuse the source ensembles, folds, and test sets of Table 3. Δ is the paired accuracy difference averaged over the six transfers, with 95% t-intervals over ten repetitions.

W1. Novelty. We agree that the backbone modules are standard, but how to design the main model architecture to effectively extract sleep signals is a meaningful engineering problem.. Each backbone component is ablated in Table 5, e.g., removing the cross-epoch Sequence module lowers N1 F1 by 6.0 points, consistent with Proposition 1. Our main methodological contributions are the two adaptation components. Recording-wise statistics alignment corrects BN statistics per recording relative to source reference moments and reweights stages to separate acquisition shift from stage composition. TypCert selects target subjects from unlabeled predictions and matches the update scope to the label budget. The ablations below show that each part contributes beyond DSA, BN re-estimation, and random selection.

W2. Complexity. On one RTX 6000 Ada GPU (one night = 1,000 epochs):

ModelInputParams (M)GFLOPs/epochms/night
DeepSleepNet1 EEG22.60.1024
TinySleepNet1 EEG1.50.065
SleepDGEEG + EOG3.40.279
CtxSleepNet (single)2 EEG + EOG15.63.64192
CtxSleepNet (10 models)2 EEG + EOG155.836.41,961
+ statistics alignment2 EEG + EOG155.847.33,446

CtxSleepNet has fewer parameters than DeepSleepNet but more FLOPs, almost all in the within-epoch encoder (CNN 1.2, GRU 0.8, Transformer 1.6 GFLOPs), while the cross-epoch LSTM adds less than 0.01. A whole night is still scored in 0.2 s by one model and in 3.4 s by the ensemble with alignment. The ensemble members are ten checkpoints of one training run, so it adds no training cost, and where inference cost matters a single model keeps most of the accuracy (85.9% vs. 87.3% on SleepEDF-20, Table 5).

Q. Stage reweighting and recording-wise BN correction. Zero-shot, without parameter updates:

BN statisticsAccMF1ΔAcc vs. no adaptation
Stored source statistics (no adaptation)68.264.4
DSA of Fan et al. (2022): one target-wide set67.964.3−0.4 [−0.8, 0.1]
Recording-wise BN re-estimation71.868.8+3.6 [2.5, 4.7]
Ours: stage reweighting only (Eq. 1, on outputs)66.062.2−2.2 [−2.5, −1.9]
Ours: BN correction relative to source only72.970.3+4.7 [3.6, 5.8]
Ours: both (full method)74.471.5+6.1 [5.3, 6.9]

The recording-wise BN correction carries most of the gain. It exceeds DSA by +5.0 [4.2, 5.9] and plain re-estimation by +1.1 [0.9, 1.3], because the source reference makes the robust geometric-mean variance usable (its scale differs from the stored variance). Stage reweighting adds +1.4 [0.9, 1.9] inside the statistics, up to +3.9 on Phys→EDF. Without the BN correction it can only rescale the outputs as a prior correction, which lowers accuracy in all six transfers because the self-estimated prior amplifies the model's own bias. DSA pools one set of statistics over all target recordings. This helps on SleepEDF targets (+6.1, +3.2), whose recordings have similar amplitudes (across-recording CV 0.24), but not on ISRUC and Physio2018 targets (−7.0 to +0.8), whose amplitudes vary far more (CV 0.80 and 0.63), so the pooled statistics fit no single recording. With fine-tuning on identical cohorts and updates, reweighting changes accuracy by +1.6 [1.5, 1.8] at 1–2%, so it helps when labels are absent or scarce.

Q. Typicality vs. certainty. At 1%/2% budgets with the same update scope (normalization + classifier):

Selection1%2%Runs < zero-shot (1%/2%, of 60)
Random69.169.418 / 13
Typicality only70.471.613 / 7
Certainty only68.870.922 / 10
TypCert70.971.99 / 2

TypCert exceeds random by +1.8 [−0.1, 3.7] and +2.5 [0.4, 4.7] (EDF→ISRUC +7.8 and +8.4), certainty alone by +2.1 [0.9, 3.2] and +1.0 [0.6, 1.4]. Certainty alone can fall below random because it favors subjects the source model already finds easy. Their cohorts contain fewer N1 epochs (9% vs. 13% for random) and match the target stage distribution worst (true TV 0.18 vs. 0.15 for random at 1%), which biases the classifier update. We keep certainty because it complements typicality. Lemma B.1 bounds the true cohort–target mismatch by dTV+Hˉd_{TV}+\bar H, and typicality alone controls only the first term. TypCert thus matches typicality in stage composition (true TV 0.10 and 0.07 at 1% and 2%) with lower predictive entropy (0.46 vs. 0.56). It also has the fewest runs below zero-shot and the lowest across-repetition SD (2.3 vs. 3.0 and 3.5 at 1%). We will add the complexity table and both ablations to the revision.

Thank you for the clarification and additional experiments. While the new results address my concerns about ablations and computational complexity, my concern regarding limited methodological novelty remains. I will maintain my original rating of 2 (Weak Reject).

We thank the reviewer for the follow-up and are glad that the new ablations and the complexity analysis resolved those concerns. Our first reply focused on experiments, so we would like to state the methodological contribution and its theoretical basis more precisely.

What is new. The backbone modules are standard and are not our main claim. The contribution is a budget-aware way to adapt a sleep-staging model to a new dataset, with one recipe from zero to 10% labels instead of a separate tool per regime:

ComponentClosest prior workWhat is newEvidence
Recording-wise BN correctionAdaBN, DSAPer-recording correction relative to source reference moments+6.5 vs. DSA
Stage-aware normalizationPrior correction of outputsStage weights inside the BN moments+1.4 at 0%, +1.6 at 1–2%
TypCertRandom or active-learning samplingLabel-free subject selection targeting dTV+Hˉd_{TV}+\bar HFewest runs below zero shot
Budget-dependent scopeOne scope for all budgetsScope set by meffm_{\rm eff}+8.4 / +8.0 on ISRUC targets

Theoretical foundation. Each design choice follows from a result in Section 3, and each prediction is confirmed by the experiments.

Proposition 1 (value of cross-epoch context). Using neighboring epochs lowers the optimal weighted risk by exactly E[w(Yi)] Iw(Yi;Z−i∣Zi)≥0\mathbb E[w(Y_i)]\,I_w(Y_i;Z_{-i}\mid Z_i)\geq 0, the label information that the context carries beyond the epoch itself, and Appendix B.1 shows that a learned sequence encoder realizes this gain once it exceeds its compression and fitting errors. The gain is large when a stage is ambiguous within one epoch, as N1 is, which motivates separating within-epoch encoding from cross-epoch modeling. Consistently, the Sequence module adds 6.0 points of N1 F1 (Table 5).

Theorem 1 (cohort bias and adaptation capacity). The error of the labeled-cohort gradient decomposes exactly into three terms: the target signal outside the trainable subspace, ∥(I−P)gU∥2\|(I-P)g_U\|^2, the cohort bias, bounded by BA=2G[dTV(A)+Hˉ(A)+εA]+ρAB_A=2G[d_{TV}(A)+\bar H(A)+\varepsilon_A]+\rho_A, and the variance, bounded by σ2rank(P)/meff(A)\sigma^2\mathrm{rank}(P)/m_{\rm eff}(A). The local risk-decrease corollary in Appendix B adds that an update lowers target risk whenever the target gradient in the trainable subspace exceeds bias plus variance. This yields the two label-dependent designs.

  • TypCert targets the bias terms. Lemma B.1 bounds the true cohort–target stage mismatch by dTV+Hˉd_{TV}+\bar H plus posterior-error terms, and each greedy step of the rank rule picks a candidate that no other candidate improves on both terms. Empirically, TypCert attains the smallest dTV+Hˉd_{TV}+\bar H of all rules at 1% (0.55 vs. 0.63 for typicality, 0.60 for certainty, and 0.83 for random) and the fewest runs below zero shot (9 and 2 of 60 vs. 18 and 13 for random).
  • The update scope follows the variance term. Updating the encoder pays off only when its additional target signal exceeds the added bias and variance, and the variance bound σ2rank(Q)/meff\sigma^2\mathrm{rank}(Q)/m_{\rm eff} is large when a cohort contains one or two subjects. Hence normalization and classifier updates at 1–2%, which with TypCert exceed random full fine-tuning by 8.4 and 8.0 points on ISRUC targets, and full adaptation at 5–10%, which reaches 79.6% at 10%, 2.1 points below in-domain training.

Why it matters. Without labels or parameter updates, mean accuracy over six transfers rises from 68.2% to 74.4%. The adaptation is not tied to our backbone and adds +7.2 [6.6, 7.8] points to a backbone trained with the SleepDG objective. With the same inputs, splits, and ensemble, CtxSleepNet also exceeds all six Table 2 baselines retrained with our pipeline (87.3% vs. at most 85.6%, see our reply to Reviewer THRu).

We will rewrite the contribution list in the Introduction around these points, add the table above to Section 2, and add a paragraph to Section 3 that maps each theoretical term to its design choice and empirical test. We hope this clarifies the methodological contribution and would be grateful if the reviewer would take it into account in the final assessment.

Dear Reviewer tY6P,

Thank you again for your thoughtful review and follow-up. Your comments have greatly improved the quality of our paper, leading to the ablations of the adaptation strategy, the complexity comparison, and a clearer statement of our methodological contribution and its theoretical basis. As the author-reviewer discussion period ends soon, we would greatly appreciate knowing whether our latest reply has addressed your concern about methodological novelty, and we look forward to your final evaluation.

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

This paper proposes CtxSleepNet for multimodal sleep staging and cross-dataset adaptation with limited target labels. The framework combines a hierarchical CNN–GRU–Transformer encoder with cross-epoch LSTM modeling, recording-wise statistics alignment, and a subject-selection strategy called TypCert. The experiments report improvements across several datasets and transfer settings, with the largest gains coming from target-domain statistical adaptation.

Critical strengths

The problem is practically relevant. Generalization across recording configurations and clinical datasets is important for sleep staging. Evaluating both unlabeled adaptation and limited-label fine-tuning makes the study useful beyond standard in-domain evaluation.

The reported adaptation gains are encouraging. Recording-wise alignment improves mean accuracy from 68.2% to 74.4% across six transfer directions. TypCert also reduces the severe performance degradation and variability observed in some low-budget settings.

Stage-aware normalization has a reasonable motivation. Feature statistics may change because of either acquisition conditions or differences in sleep-stage composition. Reweighting stages before estimating normalization statistics is a sensible attempt to separate these effects.

Critical weaknesses

Limited methodological novelty. The backbone combines standard components, and hierarchical intra-/inter-epoch modeling is already established in prior work. The theory provides general motivation but offers limited support for the specific design choices.

Insufficient adaptation ablations. The individual contributions of stage reweighting, source-reference correction, typicality, and certainty remain unclear. Comparisons with standard BN recalibration and DSA, while separately controlling subject selection and update scope, are needed.

Unclear comparison fairness. The ten-model ensemble contributes 1.4 points on SleepEDF-20, exceeding the reported 1.0-point advantage over XSleepNet. Matching input channels, ensemble settings, and target-data access across methods is necessary to establish a fair comparison.

Questions for authors
  • Can you reconcile the ISRUC baseline scores in Table 6 with the corrected dataset mapping in Wang et al. (2024), where several values correspond to HMC, and clarify how this affects the claimed improvements?

  • Under matched adaptation settings, what evidence isolates the benefit of TypCert over random selection and of the proposed normalization over standard BN recalibration or DSA, independently of the parameter-update scope?

  • Were the Table 2 baselines evaluated with matching input channels, subject splits, and ensemble settings, and do the reported advantages persist under matched single-model comparisons?

1 comment

We thank the reviewer for the careful review. All new experiments reuse the source ensembles, folds, and test sets of Table 3. Δ is the paired accuracy difference averaged over the six transfers, with 95% t-intervals over ten repetitions.

W1. Novelty and theory. We agree that the backbone modules and the intra-/inter-epoch hierarchy are established. Our contribution is the adaptation design, and (b)–(c) test its theoretical predictions. Theorem 1 splits the adaptation-gradient error into cohort bias, which TypCert targets via dTVd_{TV} and Hˉ\bar H, and variance σ2rank(P)/meff\sigma^2\mathrm{rank}(P)/m_{\rm eff}, which the restricted scope targets. As predicted, selection helps with the scope fixed, mainly on ISRUC targets. Lemma B.1 bounds the true cohort–target mismatch by dTV+Hˉd_{TV}+\bar H, which is why certainty fails alone but helps within TypCert, cutting runs below zero-shot from 13 to 9 (1%) and from 7 to 2 (2%). The restricted scope pays off only when meffm_{\rm eff} is small, so the scope should follow the number of labeled subjects (to be stated in the revision).

W2/Q2. Isolating normalization, selection, and update scope.

(a) Normalization, without selection or updates (zero-shot):

BN statisticsAccΔAcc vs. no adaptation
Stored source statistics (no adaptation)68.2
DSA of Fan et al. (2022): one target-wide set67.9−0.4 [−0.8, 0.1]
Recording-wise BN re-estimation71.8+3.6 [2.5, 4.7]
Ours: stage reweighting only (Eq. 1, on outputs)66.0−2.2 [−2.5, −1.9]
Ours: BN correction relative to source only72.9+4.7 [3.6, 5.8]
Ours: both (full method)74.4+6.1 [5.3, 6.9]

Our normalization exceeds DSA by +6.5 [5.8, 7.1] and standard recalibration by +2.5 [1.9, 3.0]. The source reference alone does not help (−0.6) but enables the robust geometric-mean variance (+1.6), whose scale differs from the stored one. Stage reweighting adds +1.4 [0.9, 1.9] inside the statistics but hurts on the outputs alone (−2.2), where the self-estimated prior amplifies the model's bias. With fine-tuning, it changes accuracy by +1.6 [1.5, 1.8] at 1–2%. DSA pools one set of statistics over all target recordings, which helps on SleepEDF targets (+6.1, +3.2) but not on ISRUC and Physio2018 (−7.0 to +0.8), whose recording amplitudes vary far more (CV 0.80 and 0.63 vs. 0.24), so pooled statistics fit no single recording.

(b) Selection, with the update scope fixed (normalization + classifier):

Selection1%2%Runs < zero-shot (1%/2%, of 60)
Random69.169.418 / 13
Typicality only70.471.613 / 7
Certainty only68.870.922 / 10
TypCert70.971.99 / 2

TypCert exceeds random by +1.8 [−0.1, 3.7] and +2.5 [0.4, 4.7] (1%/2%, EDF→ISRUC +7.8/+8.4), certainty alone by +2.1 [0.9, 3.2] and +1.0 [0.6, 1.4], and typicality alone by +0.5 [−0.8, 1.8] and +0.4 [−0.2, 0.9]. Certainty alone favors subjects the model already finds easy, with fewer N1 epochs and the worst match to the target stage distribution (true TV 0.18 vs. 0.15 for random at 1%), so it can fall below random.

(c) Update scope, with selection fixed (random cohorts): updating only normalization + classifier changes accuracy by +1.8 [−0.1, 3.6] at 1% and +0.1 [−0.9, 1.1] at 2%. At 1%, it helps on ISRUC and SleepEDF targets, where the cohort is one subject, but not on Physio2018, where it has about ten.

W3/Q3. Comparison fairness. Table 2 baselines are cited from published results: single models, mostly one EEG channel, and subject-wise CV (LOSO on SleepEDF-20, as ours). We agree that the ensemble matters at this margin. As a single model (Table 5, Case 6), it obtains 85.9/80.8/80.9 (Acc/MF1/κ) on SleepEDF-20 vs. 86.3/80.6/81.3 for XSleepNet. The accuracy margin thus relies on the ensemble, whereas the highest N1 F1 in Table 2 persists (54.1 vs. ≤52.2). We will list each method's channels and ensemble in Table 2 and adjust the in-domain claims. Across datasets, we matched data, channels, splits, ensemble, and target access by training our backbone with SleepDG's objective (source subjects as domains). Without target data, SleepDG trails CtxSleepNet by 2.2 [1.6, 2.7]. With the same unlabeled target access, our alignment raises SleepDG by 7.2 [6.6, 7.8], and CtxSleepNet stays ahead by 1.1 [0.7, 1.4] in five of six transfers.

Q1. Table 6. Thank you for catching this. All nine ISRUC entries in Table 6 come from column II of Wang et al. (2024b), which the corrected mapping (arXiv:2401.05363v5) assigns to HMC. Corrected ISRUC accuracies are 69.2–70.6% for the five networks and 72.7–75.1% for MMD, CORAL, IRM, and DANN. Trained on one labeled source rather than four, CtxSleepNet (71.8% and 71.7%) exceeds these networks but trails the DG methods. We will update Table 6 and its discussion. The Physio2018 and SleepEDF-78 entries and the main-text results are unaffected.

Reviewer_3JNpOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to adapt a multimodal sleep staging model across datasets when labels are limited. CtxSleepNet has two parts: a CNN, GRU and Transformer encoder that reads each epoch and a bidirectional LSTM that links neighbouring epochs. TYPCERT selects target subjects by typicality and prediction confidence and the labeling budget decides how many parameters are updated. When target labels are missing, adaptation reweights sleep stages and corrects normalization statistics for each recording. Tests inside one domain use SleepEDF 20, SleepEDF 78, ISRUC and Physio2018; tests that transfer across datasets use the last three. Without target labels, statistical adaptation lifts mean accuracy from 68.2% to 74.4% over six transfer directions. With a 10% labeling budget, mean accuracy reaches 79.6%.

Critical strengths

1.The experiments test several labeling settings that matter for real transfer. The authors compare adaptation with no target labels and with a few target labels across six transfer directions. The few label runs repeat splits by subject ten times and test on subjects not seen in training. Page 7 leaves out transfer between SleepEDF-20 and SleepEDF-78 because these sets share subjects. Table 3 shows how much performance drops when fine tuning uses very few labels and how much it gains back after sample selection. This makes the value of a small labeling budget clear. 2.Ablations support the architectural contributions fairly well. Table 5 on page 8 tests the GRU, the Transformer, the module that links epochs and the ensemble. Removing the epoch linking module lowers N1 F1 from 54.5 to 48.5. Appendix B explains epoch linking and adaptation under low labeling budgets using conditional mutual information and local gradient error and states when the conclusions hold.

Critical weaknesses

1.The adaptation designs need more focused validation. Table 3 on page 7 compares TYPCERT with random selection but does not test typicality and predictive certainty separately so it is unclear whether the two criteria work better together than alone. Figure 2 compares full statistical adaptation with none while the page 4 method adds stage reweighting, geometric mean variance estimation and correction against source statistics all at once. Gains over a simple BN statistics update or the cited DSA method are not shown. Please add these comparisons with the same model, data and update strategy to show which changes drive the gains. 2.Comparisons with existing methods do not yet show a clear advantage. Table 6 on page 13 lists the training sources and the UDA and DG settings and pages 13 to 14 show that CtxSleepNet can use unlabeled target data. The accuracy gaps still reflect different training sources and different access to target data so the results only serve as a reference across settings. Please compare adaptation methods with the same data input channels and target data access. Within the same domain, the origin of the results and the ensemble settings in Table 2 also need clarifying. On SleepEDF-20, Table 5 reports 85.9% for a single model and 87.3% for the full ensemble so small advantages there should be read with the ensemble in mind.

Questions for authors
  1. Table 3 reports random selection and TYPCERT results. Please add results for typicality only and for predictive certainty only with the same labeling budgets and update settings so the gain from combining both criteria is visible.
  2. With the same backbone, the same source and target splits and the same access to target data, how much does full statistical adaptation improve over a simple BN statistics update and the cited DSA method? Please first run an ablation without stage reweighting to test the part designed for different sleep stage proportions.
  3. Please clarify whether the baseline results in Table 2 come from the original papers or from rerunning the methods and give their input channels and ensemble settings. The paper says the limited label experiments were repeated ten times. Do the ± values in Table 3 show standard deviations over these splits? Please use these repeats to give paired confidence intervals for the main differences between methods.
  4. Table 6 follows the multi source setting of Wang et al. and includes baselines such as MMD, CORAL, IRM, and DANN but not SleepDG, the method from that work. Please explain why these baselines were chosen and whether SleepDG was considered.An explanation of the differences would be enough.
  5. In Table 4 on page 8, the T=20 row reports an N1 F1 of 54.9 while the matching baseline results in Table 2 on page 6 and Table 5 on page 8 report 54.5. Please explain the difference and correct it.
3 comments

We thank the reviewer for the precise and actionable suggestions. All new experiments reuse the source ensembles, folds, and test sets of Table 3. Δ is the paired accuracy difference averaged over the six transfers, with 95% t-intervals over ten repetitions.

Q1/W1. Typicality vs. certainty (same budgets and update scope, normalization + classifier):

Selection1%2%Runs < zero-shot (1%/2%, of 60)
Random69.169.418 / 13
Typicality only70.471.613 / 7
Certainty only68.870.922 / 10
TypCert70.971.99 / 2

TypCert beats certainty alone by +2.1 [0.9, 3.2] (1%) and +1.0 [0.6, 1.4] (2%) and random by +1.8 [−0.1, 3.7] and +2.5 [0.4, 4.7]. Certainty alone favors subjects the model already finds easy, whose cohorts have fewer N1 epochs (9% vs. 13%) and the worst match to the target stage mix (true TV 0.18 vs. 0.15 for random at 1%), so it can fall below random. Within TypCert it is useful because Lemma B.1 bounds the true cohort–target mismatch by dTV+Hˉd_{TV}+\bar H, and typicality controls only the first term. TypCert keeps the stage match of typicality (true TV 0.10, 0.07) at lower entropy and has the fewest runs below zero-shot.

Q2/W1. Full adaptation vs. BN re-estimation and DSA (zero-shot, same backbone, splits, and target access):

BN statisticsAccΔAcc vs. no adaptation
Stored source statistics (no adaptation)68.2—
DSA of Fan et al. (2022): one target-wide set67.9−0.4 [−0.8, 0.1]
Recording-wise BN re-estimation71.8+3.6 [2.5, 4.7]
Ours: stage reweighting only (Eq. 1, on outputs)66.0−2.2 [−2.5, −1.9]
Ours: BN correction relative to source only72.9+4.7 [3.6, 5.8]
Ours: both (full method)74.4+6.1 [5.3, 6.9]

Full adaptation improves on simple BN re-estimation by +2.5 [1.9, 3.0] and on DSA by +6.5 [5.8, 7.1]. Without stage reweighting (BN correction only), accuracy drops by 1.4 [0.9, 1.9] at zero shot and by 1.6 [1.5, 1.8] at 1–2%, so reweighting matters when labels are absent or scarce. On the outputs alone, reweighting lowers accuracy (−2.2), as the self-estimated prior amplifies the model's bias. The source reference and geometric-mean variance jointly add +1.1 [0.9, 1.3] over re-estimation. DSA pools one set of statistics over all target recordings, which helps on SleepEDF targets (+6.1, +3.2) but not on ISRUC and Physio2018 (−7.0 to +0.8), whose recording amplitudes vary far more (CV 0.80 and 0.63 vs. 0.24), so pooled statistics fit no single recording.

Q3/W2. Tables 2 and 3. All Table 2 baselines are cited from published papers and are single models with one EEG channel, except SalientSleepNet and SleepWaveNet (EEG + EOG). CtxSleepNet uses two EEG channels and EOG with a majority vote over ten checkpoints of one run. We agree the ensemble matters. As a single model, it reaches 85.9/80.8/80.9 (Acc/MF1/κ) on SleepEDF-20 vs. 86.3/80.6/81.3 for XSleepNet and keeps the highest N1 F1 in Table 2 (54.1 vs. ≤52.2). We will list these settings in Table 2. Yes, Table 3's ± values are SDs over the ten repetitions. In repetition k, fold k of ten disjoint subject folds supplies the labeled cohorts and the other 90% the shared test set. Paired 95% t-intervals for TypCert over random fine-tuning (Table 3) are +8.4 [2.9, 14.0] and +8.0 [0.4, 15.6] on ISRUC targets at 1% and 2%, and +3.6 [1.0, 6.2] and +2.6 [0.0, 5.2] over six transfers. For domain adaptation:

BudgetZero-shot1%2%5%10%
With vs. without adaptation+6.1 [5.3, 6.9]+6.1 [4.8, 7.3]+5.8 [5.2, 6.3]+2.3 [1.4, 3.3]+1.6 [0.7, 2.4]

Q4/w2. Matched comparison with existing methods. Table 6 cites the label-free benchmark of Wang et al. (2024b), whose multi-source methods, SleepDG included, use four labeled cohorts, no target data, and one EEG plus one EOG channel. We agree a matched comparison is needed and trained our backbone with its official objective (source subjects as domains) on our data, channels, splits, and ensemble:

Zero-shot AccEDF→ISRUCEDF→PhysISRUC→EDFISRUC→PhysPhys→EDFPhys→ISRUC
SleepDG61.071.457.671.970.563.8
CtxSleepNet62.969.567.173.571.564.7
SleepDG + our alignment70.874.473.674.074.971.8
CtxSleepNet + our alignment71.875.076.174.577.071.7

Without target data, SleepDG trails CtxSleepNet by 2.2 [1.6, 2.7]. With the same target access, our alignment raises SleepDG by 7.2 [6.6, 7.8], and CtxSleepNet stays ahead by 1.1 [0.7, 1.4] in five of six transfers, so the advantage is not due to target access. Table 6 will be marked as a cross-setting reference with corrected ISRUC entries.

Q5. Thank you. 54.9 is a typo for 54.5 and will be corrected.

Dear Reviewer 3JNp,

Thank you again for your precise and constructive review. Your suggestions have greatly improved the quality of our paper, leading to the separate evaluation of typicality and certainty, the comparison with BN re-estimation and DSA, and the matched comparisons with SleepDG and with the Table 2 baselines (see also our latest reply to Reviewer THRu). Since the author-reviewer discussion period closes soon, we would be grateful to know whether our responses have addressed your concerns, and we look forward to your final evaluation.

After careful review, I remain unconvinced of the method’s novelty and find the external validation insufficient, so I maintain my recommendation of weak rejection.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response and summary of revisions

We thank all reviewers for their careful and constructive reviews, which have greatly improved the quality of our paper. We have uploaded a revised PDF in which all changes are shown in purple. The main text stays within ten pages, and the new results are as follows.

  1. Ablation of the adaptation strategies (new Section 4.5, Figure 3, Appendices A.3–A.4). Under the same ensembles, folds, and test sets, our recording-wise alignment exceeds DSA by 6.5 points and recording-wise BN re-estimation by 2.5 points at zero shot. The BN correction relative to source references carries most of the gain. Stage reweighting adds 1.4 points at zero shot and 1.6 points at 1–2% budgets.
  2. Typicality vs. certainty (Section 4.5, Appendix A.5). With the same update scope, TypCert exceeds random, typicality-only, and certainty-only selection and has the fewest runs below zero-shot accuracy (9 and 2 of 60 at 1% and 2%). A cohort analysis explains why certainty alone can fall below random selection.
  3. Matched comparison with SleepDG (Section 4.5, Appendix A.2). With the official SleepDG objective trained on our data, channels, splits, and ensemble, CtxSleepNet exceeds SleepDG by 2.2 points without target data, and our alignment improves SleepDG by 7.2 points.
  4. Corrected and annotated baselines. Table 6 now follows the corrected dataset numbering of Wang et al. (2024b). Table 2 lists each method's input channels and our single-model result. Two values in Table 3 are corrected.
  5. Complexity (Appendix A.6). Parameters, FLOPs, and inference time of CtxSleepNet and the baselines.

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

Related papers

Discussion (0)

Sign in to comment.