acceptodds
Under review as a conference paper at ICLR 2027

SleepXA: Composing Pretrained Neural and Cardiac Foundation Models for Sleep Representation Learning

Abstract

Physiological signals recorded during sleep provide synchronized views of neural, cardiac, respiratory, and other bodily systems, while preserving modality-specific information. Existing EEG and ECG foundation models already capture rich but distinct physiological structure, motivating an alternative to training a multimodal sleep model from scratch: composing existing pretrained models. We introduce SleepXA, a framework that uses synchronized polysomnography (PSG) to compose independently pretrained neural and cardiac foundation models. SleepXA combines pretrained EEG and ECG encoders with lightweight encoders for additional PSG modalities, and learns shared and modality-specific representations through sleep-guided alignment. We pretrain SleepXA on 9,230 overnight PSG recordings from three large cohorts and evaluate it on 27 downstream tasks spanning sleep representation, clinical and disease prediction, and daytime physiological transfer. SleepXA ranks among the top two methods on 22 of 27 tasks. Controlled experiments further show consistent benefits across sensor configurations and EEG–ECG backbone choices, while pretrained initialization outperforms architecture-matched training from scratch and SleepXA surpasses alternative multimodal composition strategies. These results suggest that synchronized multimodal physiology can serve as a natural interface for composing existing physiological foundation models, yielding reusable representations that extend beyond sleep to clinical and daytime physiological settings.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Sleep Representation Learning, Physiological Time Series, Foundation Models for Health

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_KjKFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SleepXA uses pretrained models for EEG and ECG along with a small network for remaining channels to create a sleep foundation model applicable to polysomnography data. The model was trained on thee datasets: SHHS, MESA and CFS totalling 9230 PSGs. The model is trained in stages unfreesing layers slowly aligning the representation of EEG, EKG and remainder. The modalities are aligned into a shared representation, and sleep-stage labels are used during pretraining. The authors evaluate the model on 27 tasks including sleep staging sleep apnea, disease outcomes and non sleep tasks. The paper reports top-2 on 22 of 27 tasks.

Critical strengths
  1. Good set of tasks (27 diverse tasks including canonical sleep, disease, an other biosignal tasks) on a large pretraining dataset. The model is pretrained on 9,230 PSGs (a decent size although it could be made a lot bigger, specially for pretraining purposes) from three cohorts and evaluated on several external cohorts, with both linear probing and fine-tuning.

  2. The paper aims to solve an important problem with modelling different modalities into a shared space based on occuring at the same time by applying losses that allow the model to disentangle cohort, sleep stage and modality from this main loss.

The manuscript is overall well written and the figures a

Critical weaknesses
  1. reliability of the evaluations (overstating performance):
  • Some of the SleepFM results are so far off what is reported in the SleepFM paper itself (https://www.nature.com/articles/s41591-025-04133-4 fig3) compared to Table A.10 composit scores.
  • The sleep staging label creeps in at pretraining time esdsnetially allowing the representations of the proposed SleepXA to be more aligned with the task and making these results less reliable.
  1. Unclear the impact of the different loss components
  • The InfoNCE alignment is a variant of SleepFM's objective. CLIP-style alignment trails SleepXA by only 0.2–1.2 MF1 under LP (Table A.14).
  • Therefore its unclear if the added losses actually help meaningfully on pretraining
Questions for authors

Could you add a clear analysis on how the different losses contribute?

Why is your reported SleepFM results underperforming the published values so much?

What is the effect on the sleep staging performance based on the sleep stage-based loss?

8 comments
AuthorsOpenReview ↗

Response to Reviewer KjKF (Part 1)

We are grateful for the reviewer's attentive analysis and helpful feedback.

W1 / Q1: Contributions of the learning objectives

W1: The contributions of the additional loss components are unclear, particularly given the small LP gains over CLIP-style alignment in Table A.14.

Q1: Could the authors provide a clear analysis of how the individual loss components contribute to pretraining and downstream performance?

Response.

We thank the reviewer for this question. The additional objectives implement sleep-guided cross-modality common–specific learning for composing independently pretrained physiological encoders. They serve complementary roles: common-space alignment encourages cross-modal correspondence within matched cohort–stage conditions; stage-conditioned specific learning preserves complementary modality-dependent information; and conditional residual disentanglement penalizes redundancy after removing condition-specific means. Reconstruction additionally encourages information preservation.

The complete objective achieves the highest LP F1 on all three evaluated tasks: WSC sleep staging, WSC OSA severity classification, and CinC2017 AF recognition (Tables A.18–A.20). The LP gain over CLIP-style alignment is 0.83 percentage points for WSC sleep staging (Table A.16, formerly Table A.14) and 19.75 points for AF recognition (Table A.17; Table R16 below). Thus, sleep-staging performance alone does not fully reflect the benefit of the combined objectives for cross-context transfer. This pattern is also consistent with the stage-guidance experiments, whose contribution we discuss separately in our response to W3.

Table R16. Cross-context transfer on the CinC2017 AF task. AF-F1 scores (%) are reported under linear probing (LP).

MethodLP AF-F1 (%)
SleepXA (Ours)74.30
MindMix-style Alignment51.83
Cross-Attention Fusion61.38
Concatenation + MLP Fusion49.79
DecAlign-style Alignment47.02
CLIP-style Contrastive Alignment54.55
AuthorsOpenReview ↗

Response to Reviewer KjKF (Part 2)

W2 / Q2: Clarification of the SleepFM evaluation protocol

W2: The reported SleepFM results are substantially lower than its published results, raising concerns about baseline evaluation and the reliability of the performance comparisons.

Q2: Why does SleepFM underperform its published results by such a large margin in your evaluation?

Response.

Thank you for highlighting this discrepancy. We recognize SleepFM as a pioneering contribution to sleep foundation modeling and disease prediction. Our reported results are a standardized re-evaluation of SleepFM, not a reproduction of its original Nature Medicine experiment[1]. We would like to clarify an important distinction between our evaluation protocol and that used in the original study.

In our study, SHHS contributes to PSG alignment, while disease evaluation uses a separate paired SHHS1–SHHS2 cohort comprising approximately 1,000 participants in total. We use pooled frozen features and a linear prediction head without demographic covariates. All compared models are evaluated using this same smaller downstream cohort and protocol.

By comparison, SleepFM was pretrained on large external cohorts, including the Stanford Sleep Clinic cohort, allowing SHHS to be reserved for downstream adaptation and evaluation. Its original SHHS evaluation used 3,291 training participants and LSTM-based prediction with age and sex [1]. Our study does not have access to a comparable external clinical pretraining cohort. The available downstream training data, prediction heads, covariates, outcome definitions, and evaluation metrics therefore differ. These differences provide important context for the relatively modest disease-prediction scores observed across models in our benchmark.

SleepMaMi provides direct evidence of the effect of the probing protocol: SleepFM’s staging MF1 increases from 56.3% under linear probing to 78.0% with an LSTM classifier [2]. We conducted a similar evaluation on Test SHHS1/2, where LSTM probing also improved SleepFM’s sleep-staging performance (Table R17).

Table R17. Effect of the probing protocol on Test SHHS1/2 sleep staging.

Model / probing protocolAcc. (%)MF1 (%)κ
SleepFM75.2268.050.667
SleepFM + LSTM83.6776.210.772
SleepFM + LSTM + age/sex83.0076.410.766
SleepXA84.0776.240.780
SleepXA + LSTM86.1979.900.807

For disease prediction, using an LSTM head with age and sex improves performance, although scores remain modest, potentially reflecting the smaller cohort and limited case numbers.

Recent independent studies, including sleep2vec (ICLR 2026), OSF (ICML 2026), Omni-Sleep (MICCAI 2026), and SleepMaMi (ICML 2026), have also reported lower SleepFM scores in some downstream settings than those reported in the original Nature Medicine study [2–5]. These findings are broadly consistent with our observations, and the performance differences may partly reflect variations in datasets and evaluation protocols.

We will clarify these distinctions in the revised manuscript, and open-source our experimental settings for fair benchmarking.

References

[1] R. Thapa et al. A multimodal sleep foundation model for disease prediction. Nature Medicine, 32:752–762, 2026.

[2] K. Park, Y. Na, Y. R. Choi, H. Ryu, H.-W. Shin, and H.-S. Kim. SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structures. ICML, 2026.

[3] W. Yuan, Z. Jin, Y. Wang, D. Xie, Z. Ye, C. Zhang, and X. Chen. sleep2vec: Unified Cross-Modal Alignment for Heterogeneous Nocturnal Biosignals. ICLR, 2026.

[4] Z. Shuai, Z. Xu, D. Yang, W. Wang, and Y. Yang. OSF: On Pre-training and Scaling of Sleep Foundation Models. ICML, 2026.

[5] Z. Hou, S. Wang, K. Lou, M. Wang, C. Wei, and Q. Liu. Omni-Sleep: A Sleep Foundation Model via Hierarchical Contrastive Learning of CNS–ANS Dynamics. MICCAI, pp. 439–448, 2026.

AuthorsOpenReview ↗

Response to Reviewer KjKF (Part 3)

W3 / Q3: Contribution of sleep-stage guidance

W3: Using sleep-stage labels during pretraining may favor SleepXA on downstream sleep staging and complicate comparisons with models trained under different supervision settings.

Q3: How much does sleep-stage guidance contribute to sleep-staging performance, and how does performance change when this guidance is removed?

Response.

We thank the reviewer for raising this point. Sleep-stage guidance is an explicit feature of SleepXA, providing a physiological reference for aligning independently pretrained unimodal encoders. We appreciate the difficulty of acquiring such labels, especially training on large-scale label-free sleep recordings. Therefore, in line with our responses to the other two reviewers, we provide several alternatives reducing the reliance on the “full” sleep stage label, and therefore assess its contribution.

The experiments were conducted using four settings: (i) Full uses expert labels from SHHS, MESA, and CFS; (ii) None removes stage-dependent objectives while retaining instance contrastive learning, reconstruction, and other objectives that do not require sleep-stage labels; (iii) Hybrid uses expert SHHS labels and pseudo-labels for MESA/CFS generated by a model trained only on SHHS; and (iv) Partial uses expert SHHS labels while MESA/CFS contribute only to objectives that do not require stage labels.

As shown in Table R18, removing guidance reduces WSC staging MF1 by 1.17/1.26 percentage points and accuracy by 1.45/0.55 percentage points under LP/FT, respectively. Hybrid and partial guidance retain much of the full-label performance, with LP MF1 reaching 75.89% and 75.57%, versus 76.09% with full guidance.

Table R18. Effect of sleep-stage guidance on WSC sleep staging.

ModeGuidanceAcc. (%)MF1 (%)κ
LPNone83.7174.920.751
LPHybrid84.6975.890.762
LPPartial84.2375.570.760
LPFull (SleepXA)85.1676.090.765
FTNone85.4076.350.770
FTHybrid85.5177.190.774
FTPartial84.9076.200.765
FTFull (SleepXA)85.9577.610.780

Table R19. Effect of sleep-stage guidance on WSC AHI estimation and OSA severity classification. Acc. and MF1 evaluate OSA severity classification.

ModeGuidanceMAERMSER²Acc. (%)MF1 (%)
LPNone3.6556.5840.85373.4173.68
LPHybrid3.4896.3560.86378.3173.88
LPPartial3.5086.5560.85477.5173.73
LPFull (SleepXA)3.4416.3410.86479.1275.05
FTNone3.6216.6110.85278.7275.19
FTHybrid3.2806.4920.85780.7277.83
FTPartial3.4106.5080.85680.3177.31
FTFull (SleepXA)3.1206.0080.87882.7378.46

Table R20. Effect of sleep-stage guidance on CinC2017 AF recognition. All scores are percentages.

ModeGuidanceChallenge-F1AF-F1AF-AUROC
LPNone59.5160.2487.40
LPHybrid72.9173.3589.60
LPPartial72.6072.1288.60
LPFull (SleepXA)73.1874.3091.15
FTNone82.6981.8598.30
FTHybrid82.7181.5698.20
FTPartial82.3981.4198.10
FTFull (SleepXA)83.3582.3098.40

Removing guidance also reduces OSA classification accuracy by 5.71/4.01 percentage points under LP/FT and LP AF-F1 by 14.06 percentage points (Tables R19 and R20). The AF-F1 gap narrows to 0.45 percentage points after FT. These results suggest that stage guidance supports transferable representations beyond improving sleep staging alone. This task-dependent pattern is consistent with the smaller gap between SleepXA and CLIP-style alignment on staging and the larger advantage on AF recognition.

These results demonstrate that sleep stage labels are important, but do not necessarily need to cover the whole training set for pretraining. Although full sleep-stage guidance achieves the best performance, alternative solutions such as pseudo-label and partial-label guidance can achieve comparable performance to the full version, when expert annotations are prohibitive to acquire.

We hope these results and clarifications address your concerns and support a more positive assessment of our work. We would be happy to answer any further questions.

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

This paper proposes SleepXA, a framework for composing independently pretrained EEG and ECG foundation models using synchronized polysomnography (PSG) recordings. Rather than training a multimodal sleep foundation model from scratch, the authors aim to reuse existing physiological representations while learning how to align shared information and preserve modality-specific features. The proposed framework consists of three main components: common-space alignment, modality-specific preservation, and common-specific disentanglement. SleepXA uses pretrained EEG and ECG encoders alongside lightweight encoders for additional physiological modalities. The representations are aligned through instance-level and sleep-stage supervision, while modality-specific prototypes and conditional decorrelation objectives aim to preserve complementary information. The authors pretrain the framework on 9,230 overnight PSG recordings and evaluate it on 27 downstream tasks across 15 datasets, including sleep staging, clinical outcome prediction, and transfer to non-sleep physiological tasks. The reported results are generally strong, with SleepXA ranking first on 17 tasks and among the top two on 22.

Critical strengths

The paper investigates an interesting and practically meaningful research direction. Existing physiological foundation models are typically pretrained independently on different signal modalities, and learning how to compose them without discarding their pretrained knowledge is a worthwhile problem. I find this motivation more compelling than simply proposing another multimodal sleep model trained from scratch. The idea of using synchronized physiological recordings as an interface between independently pretrained models is appealing and potentially applicable beyond the sleep domain.

The common-specific representation design is also reasonable. The authors recognize that aligning heterogeneous modalities into a single shared representation can remove useful modality-specific information. Their attempt to preserve such information while encouraging cross-modal consistency is well motivated. I particularly appreciate the use of sleep-stage and cohort conditioning in the alignment and disentanglement objectives, rather than applying the same constraints globally.

The experimental evaluation is a major strength. The paper covers a fairly extensive set of downstream tasks, including general sleep representation, clinical prediction, longitudinal outcomes, and transfer to physiological recordings outside the sleep setting. This breadth is valuable because the main contribution is not simply improving sleep staging, but demonstrating that the composed representations remain useful across different physiological contexts.

I also appreciate the controlled experiments examining the role of pretrained initialization. Figures 6 and 7 provide evidence that the gains are not solely due to the architecture, and the comparisons across different EEG and ECG backbone combinations help support the generality of the composition approach. The comparison with alternative multimodal integration strategies in Figure 8 is particularly relevant, since it addresses whether the proposed design provides benefits beyond straightforward fusion or contrastive alignment.

Overall, the paper is well organized, and the authors have made a substantial effort to support the proposed framework experimentally.

Critical weaknesses
  1. The methodological novelty of the individual components is limited.

My main reservation concerns the novelty of the proposed learning objectives. The three components of SleepXA are largely based on established ideas in multimodal representation learning, including contrastive alignment, prototype-based supervision, shared-private representation decomposition, decorrelation, and reconstruction. While the authors adapt these techniques to physiological foundation-model composition, it is not entirely clear which aspect constitutes a fundamentally new methodological contribution. The cohort- and sleep-stage-conditioned objectives are sensible design choices, but they appear to be relatively incremental modifications of existing representation-learning techniques. I think the paper's strongest novelty lies in the overall compositional setting rather than in the specific optimization formulation. The authors could make this distinction clearer and explain more precisely what new principle the proposed framework establishes for composing independently pretrained foundation models.

  1. The common-specific decomposition is not sufficiently validated beyond downstream performance.

A central claim is that SleepXA learns shared physiological information while preserving modality-specific knowledge. However, the evidence for this decomposition is largely indirect. The ablations show that removing the proposed objectives reduces downstream performance, but this does not necessarily establish that the learned common and specific representations actually capture the intended types of information. For example, it would be useful to examine whether the common representations become less predictive of modality identity while retaining sleep-stage information, and whether the specific representations preserve physiologically meaningful modality-dependent features. Such analyses would help establish that the decomposition behaves as intended, rather than simply providing additional regularization that improves downstream accuracy. Relatedly, the reconstruction objective encourages information preservation, but does not guarantee that the shared and private branches are semantically disentangled. I would have liked to see more direct representation-level evidence supporting this interpretation.

  1. The role of sleep-stage supervision deserves more careful consideration.

Although SleepXA is presented as a framework for composing pretrained physiological foundation models, its common-space alignment explicitly uses sleep-stage labels and stage-conditioned prototypes during pretraining. This is an important distinction from fully self-supervised multimodal representation learning. The method benefits from structured supervision available in the PSG cohorts, and it is not clear how much of its advantage comes from the compositional framework itself versus the additional sleep-stage supervision. The comparison with existing sleep foundation models would be more convincing if the authors isolated the effect of sleep-stage labels, for example by comparing against a purely instance-level alignment variant and a version that uses the same supervision with a simpler fusion architecture. The current ablations partially address this question, but the condition-aware alignment ablation removes multiple components together. A more targeted analysis would help clarify the contribution of stage-level supervision.

  1. The evaluation protocol and computational trade-offs could be clearer.

The breadth of the evaluation is impressive, but the comparison involves models with different pretraining histories, architectures, parameter counts, and supported input modalities. SleepXA inherits representations from two independently pretrained foundation models and performs additional multimodal alignment training. Consequently, its total pretraining resources and available prior knowledge may differ substantially from those of the competing sleep foundation models. This does not invalidate the results, since reusing pretrained models is precisely the motivation of the paper. However, it complicates claims about the efficiency or superiority of the proposed approach. I would appreciate a more explicit accounting of model size, trainable parameters, alignment-training cost, and inference overhead. This would help distinguish the benefits of compositional reuse from those of simply having access to more pretrained capacity.

  1. Some of the broad transfer claims could be interpreted more cautiously.

The results on clinical and daytime physiological tasks are interesting, but several of these evaluations involve relatively small datasets. In particular, some non-sleep transfer benchmarks contain only a few dozen subjects, which can make performance estimates sensitive to the evaluation split and training configuration. The clinical prediction results also show that the improvements are not uniformly large. For several longitudinal outcomes, the difference between SleepXA and the strongest baseline is modest relative to the reported variation. I therefore think the results support the practical transferability of the representations, but the stronger interpretation that the framework learns broadly reusable multisystem physiological knowledge would benefit from additional analysis.

Questions for authors
  1. Could the authors clarify the main methodological novelty of SleepXA relative to existing shared-private multimodal representation-learning approaches, particularly DecAlign? Which component of the proposed formulation is essential to the foundation-model composition setting?

  2. Is there direct evidence that the common and modality-specific branches learn the intended decomposition? For example, can the authors quantify modality identity, sleep-stage information, or physiological attributes retained in each branch?

  3. How much of the improvement depends on sleep-stage supervision during pretraining? Could the authors provide a more targeted comparison between instance-only alignment, stage-guided alignment, and simpler supervised multimodal fusion?

  4. Could the authors report the total computational cost and parameter count of SleepXA relative to the compared sleep foundation models, including the additional multimodal alignment stage?

  5. How stable are the improvements on the smaller clinical and daytime transfer datasets? Additional repeated-seed results or statistical comparisons would help assess the robustness of the reported gains.

6 comments
AuthorsOpenReview ↗

Response to Reviewer ArTT (Part 1)

Thank you for your thoughtful review and for recognizing the breadth of our evaluation; below, we address each concern and its corresponding question.

W1 / Q1: Novelty and comparison with shared–private approaches

W1: The individual learning objectives build on established multimodal representation-learning techniques. The novelty appears stronger in the overall compositional setting than in the optimization formulation.

Q1: How does SleepXA differ from existing shared–private approaches, particularly DecAlign, and what new principle or essential mechanism does it introduce for composing independently pretrained foundation models?

Response.

Thanks for this question. We agree that shared–private decomposition adopts existing representation-learning techniques, which we did not claim as novelty per se. Instead, our key contribution lies in providing a practical route to composing different pretrained modality-specific physiological foundation models by a sleep-guided cross-modality common-specific learning.

Further to Q1, thanks for raising these important work. DecAlign uses prototype-guided optimal transport and distribution matching for visual, acoustic, and language representations. SleepXA uses sleep stage as a shared physiological reference to guide three coupled mechanisms: common-space alignment within matched cohort–stage conditions, modality-specific preservation within each stage, and residual disentanglement after removing condition-specific means. This formulation reflects a key property of PSG: different modalities observe the same sleep state but retain distinct physiological responses. Both branches can therefore contain meaningful stage-related information, motivating conditional separation while preserving complementary pretrained knowledge. We have clarified this rationale in Sections 3.2–3.4 and expanded the downstream comparisons and ablation studies in Section 4.3 of the revised manuscript. We have also added representation-level analyses in Appendix A.6 to provide quantitative evidence on the information retained in each branch, as detailed in our response to Q2.

Tables A.16 and A.17 compare SleepXA with existing integration strategies under consistent experimental settings. Under LP, SleepXA achieves a WSC sleep-staging MF1 of 76.09% and a CinC2017 AF-F1 of 74.30%, compared with 75.47% and 47.02% for DecAlign-style alignment, respectively. The corresponding LP gains over CLIP-style alignment are 0.83 and 19.75 percentage points. After FT, SleepXA retains higher scores than DecAlign-style alignment on both tasks, with gains of 1.25 and 0.69 percentage points, respectively. These results support the value of the proposed integration strategy, particularly for cross-context transfer under LP.

AuthorsOpenReview ↗

Response to Reviewer ArTT (Part 2)

W2 / Q2: Direct representation-level evidence

W2: Downstream performance and ablations provide only indirect evidence for the intended common–specific decomposition. Reconstruction alone does not guarantee semantic disentanglement.

Q2: Can the authors provide direct representation-level evidence by quantifying modality identity, sleep-stage information, and physiological attributes retained in each branch?

Response.

Thank you for this suggestion. We agree that downstream performance and reconstruction alone do not establish semantic disentanglement. Directly validating the intended decomposition remains challenging. Following your suggestion, we conducted exploratory linear probes on frozen SleepXA representations extracted before fusion, using the same WSC splits and pretrained checkpoint. Only the probes were trained.

Table R10. Exploratory linear probing of frozen SleepXA representations.

Classification

ProbeCommon branch Acc.Common branch MF1Specific branch Acc.Specific branch MF1
Modality identity89.1189.21100.00100.00
Sleep stage82.0173.2378.6567.17

Physiological regression

ProbeCommon branch R²Common branch MAESpecific branch R²Specific branch MAE
N2 EEG relative σ power0.7130.0150.7760.013
ECG heart rate0.6453.9120.8372.76

The EEG target is the ratio of 11–16 Hz power to 0.5–30 Hz power within N2 epochs. ECG heart-rate targets were derived from automatically detected R peaks after quality screening.

The common branch better supports sleep-stage decoding, whereas the specific branch better predicts modality identity and the examined physiological attributes. These results provide preliminary representation-level evidence that the two branches retain information with different emphases, consistent with their intended complementary roles in the common–specific decomposition.

In addition, the differences in UMAP have been further confirmed. This section supplements and revises Appendix A.6.

In the future, it might be great to use these “straightforward” metrics as surrogate measure for selecting and validating common–specific decomposition strategies, rather than relying solely on downstream performance. We view systematically developing these measures and using them to guide the degree of decomposition as an interesting direction for future work.

AuthorsOpenReview ↗

Response to Reviewer ArTT (Part 3)

W3 / Q3: Effect of stage guidance and integration strategy

W3: SleepXA uses sleep-stage supervision during alignmen.....

Q3: How much of the improvement comes from stage supervision.....

Response.

Thank you for this suggestion. We agree that the original condition-aware alignment ablation alone does not fully isolate the contribution of stage supervision. We therefore compare four label settings within the same SleepXA framework and summarize the relevant instance-alignment and fusion baselines below.

Contribution of stage guidance. We fix encoder initialization, architecture, alignment data, and downstream evaluation, varying only stage-label use during alignment:

  • Full (SleepXA) uses expert SHHS, MESA, and CFS labels.
  • None removes stage-dependent objectives but retains instance contrastive learning, reconstruction, and other label-independent objectives.
  • Hybrid retains expert SHHS labels and uses pseudo-labels for MESA and CFS from a SHHS-only SleepXA model.
  • Partial retains expert SHHS labels and treats MESA and CFS as unlabeled; these recordings contribute only to label-independent objectives.

Full performs best across Tables R11–R13. Relative to None, it improves staging MF1 by 1.17/1.26 percentage points and OSA accuracy by 5.71/4.01 points under LP/FT, while reducing AHI MAE. The AF-F1 gain is 14.06 points under LP and 0.45 after FT. Hybrid and Partial recover part of the LP staging gap and most of the LP AF-F1 gap, although their LP OSA MF1 gains are small. They offer alternatives when expert annotations are limited.

Table R11. Effect of sleep-stage guidance on WSC sleep staging.

ModeGuidanceAcc. (%)MF1 (%)
LPNone83.7174.92
LPHybrid84.6975.89
LPPartial84.2375.57
LPFull (SleepXA)85.1676.09
FTNone85.4076.35
FTHybrid85.5177.19
FTPartial84.9076.20
FTFull (SleepXA)85.9577.61

Table R12. Effect of sleep-stage guidance on WSC AHI estimation and OSA severity classification.

ModeGuidanceMAER²Acc. (%)MF1 (%)
LPNone3.6550.85373.4173.68
LPHybrid3.4890.86378.3173.88
LPPartial3.5080.85477.5173.73
LPFull (SleepXA)3.4410.86479.1275.05
FTNone3.6210.85278.7275.19
FTHybrid3.2800.85780.7277.83
FTPartial3.4100.85680.3177.31
FTFull (SleepXA)3.1200.87882.7378.46

Table R13. Effect of sleep-stage guidance on CinC2017 AF recognition.

ModeGuidanceAF-F1 (%)AF-AUROC (%)
LPNone60.2487.40
LPHybrid73.3589.60
LPPartial72.1288.60
LPFull (SleepXA)74.3091.15
FTNone81.8598.30
FTHybrid81.5698.20
FTPartial81.4198.10
FTFull (SleepXA)82.3098.40

Comparison with instance-only alignment and simpler fusion. Table R14 brings together the CLIP-style instance-only alignment baseline, Concatenation + MLP Fusion, and SleepXA, using the results in Tables A.16–A.17.

Under LP, SleepXA exceeds CLIP-style alignment by 0.83 percentage points in WSC staging MF1 and 19.75 points in AF-F1. Compared with Concatenation + MLP Fusion, the corresponding LP gains are 1.63 and 24.51 points. After FT, SleepXA retains higher WSC staging MF1, while AF-F1 is nearly identical to that of Concatenation + MLP Fusion (82.30% versus 82.29%).

Table R14. Comparison of instance-only alignment, simple fusion, and SleepXA.

MethodWSC staging MF1: LPWSC staging MF1: FTCinC2017 AF-F1: LPCinC2017 AF-F1: FT
CLIP-style instance-only alignment75.2675.8454.5581.94
Concatenation + MLP Fusion74.4676.5749.7982.29
SleepXA (None)74.9276.3560.2481.85
SleepXA (Full)76.0977.6174.3082.30
Reviewer_2diDOpenReview ↗
Rating
2
Confidence
4
Informativeness
3(+1)
Presentation
2
Soundness
2
Read review
Paper summary

SleepXA composes independently pretrained EEG and ECG foundation models with lightweight encoders for other PSG modalities. It uses sleep-stage-supervised objectives to align common representations, preserve modality-specific information, and regularize common–specific redundancy. The model is trained on 9,230 PSG recordings and evaluated on 27 tasks across 15 datasets, including sleep analysis, clinical prediction, and non-sleep physiological transfer.

Critical strengths
  1. Meaningful research direction. Reusing independently pretrained physiological models through synchronized PSG is an interesting approach, compared to developing another sleep foundation model entirely from scratch. The experiments with different EEG–ECG backbone combinations strengthen this contribution.

  2. Broad evaluation with useful controls. The paper includes external cohorts, non-sleep tasks, initialization studies, integration baselines, and objective ablations. Frozen-encoder linear probing provides meaningful evidence beyond end-to-end fine-tuning.

Critical weaknesses
  1. An important, closely related reference is missed: SleepMaMi [1]. SleepMaMi designs both private and shared encoders to learn modality-specific and shared representations, which makes the claim that existing sleep foundation models focus on learning common representations not true. Given that SleepMaMi is also shown to outperform SleepFM, SleepMaMi should be analyzed both qualitatively and quantitatively to make the novelty of this work clear. Some of the statements in this paper might need to be adjusted accordingly.

  2. Given that SleepMaMi learns both private and shared representations, and existing sleep foundation models already learn 20k (SleepMaMi), 65k (SleepFM), and 235k (SleepFM-2) PSG cases, it is unclear if composing separately pretrained modality-specific encoders benefits in practice. Is it practical to assume that EEG or ECG foundation models separately trained without using these large PSG datasets add significant value?

  3. While existing sleep foundation models utilize self-supervised learning, the proposed learning mechanism heavily relies on sleep stage labels. It is unclear if this sleep stage-dependent supervised learning is desirable for foundation model. (1) Scalability: Sleep stage labels need full PSG recording and labor/time cost. (2) Reliability: It is known that sleep technicians agree to others' sleep stage labels only 82%, which verifies significant label inconsistency.

[1] K. Park et al., "SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structures," ICML, July 2026.

Questions for authors
  1. What if SleepXA adopts SleepMaMi's encoder architecture and learning methods for integrating pretrained EEG and ECG models?
10 comments
AuthorsOpenReview ↗

Response to Reviewer 2diD (Part 1)

Thank you for your constructive feedback and for recognizing the value of pretrained model composition and our broad evaluation. We have added the missing comparison and further analyses to address your concerns.

W1: SleepMaMi is missing, it also learns shared and private representations. Qualitative and quantitative comparisons are needed.

Response.

Thank you for pointing out SleepMaMi [1], a closely related concurrent work published at ICML in July 2026*. We have now discussed it in the introduction and related work and included it in quantitative comparisons on the main downstream tasks. We have also revised the overly broad statement about existing sleep foundation models’ treatment of modality-specific information.

However, SleepMaMi and SleepXA are conceptually different in many ways. The key distinction lies in the utilisation of pretrained model or not. SleepMaMi learns a PSG representation hierarchy using a shared-private Micro-Encoder and a full-night Macro-Encoder, from scratch. SleepXA instead investigates how to compose independently pretrained EEG and ECG foundation models through synchronized PSG, and in turn to support more general non-sleep tasks.

In particular, composing knowledge from existing pretrained foundation models brings both opportunities (reusing rich modality-specific knowledge) and challenges (aligning heterogeneous pretrained representation spaces without losing complementary modality-specific information). SleepXA is designed to address these challenges through sleep-guided common–specific learning.

We evaluated SleepMaMi using the same data splits and LP and FT protocols as the other baseline models. Given the limited revision time, our additional experiments focused on major tasks, including sleep staging, whole-night AHI estimation, and OSA severity classification. The results are reported in Table 1 and Tables A.7–A.11 of the revised manuscript.

We list some of them here.

Table R1. Sleep staging .

CohortModelLP Acc.LP MF1FT Acc.FT MF1
Test SHHS1/2SleepFM75.2268.0574.5167.73
SleepGPT77.9169.7582.1374.37
SleepMaMi80.2869.9284.6475.13
OSF83.1474.4284.0975.58
SleepXA84.0776.2487.0478.71
APPLESSleepFM70.8962.7072.0962.95
SleepGPT72.2463.9273.0666.51
SleepMaMi75.2865.6977.4767.91
OSF75.5067.3076.7268.20
SleepXA77.6371.1377.8470.51
WSCSleepFM77.4067.9768.7363.79
SleepGPT79.5168.3582.4671.54
SleepMaMi81.3371.5185.1677.09
OSF83.5275.6183.1874.97
SleepXA85.1676.0985.9577.61

Table R2. AHI estimation (R²) and OSA severity classification (MF1).

ModeModelTest SHHS1/2 R²Test SHHS1/2 MF1WSC R²WSC MF1
LPSleepFM0.65245.980.80963.89
SleepGPT0.70553.370.83066.72
SleepMaMi0.79367.230.83970.51
OSF0.76465.340.81767.29
SleepXA0.79968.350.86475.05
FTSleepFM0.78059.210.81867.35
SleepGPT0.79468.370.89976.28
SleepMaMi0.80970.510.86874.95
OSF0.81460.610.88072.69
SleepXA0.84273.320.87878.46

SleepMaMi is a strong comparator, particularly for whole-night AHI estimation. SleepXA achieves higher point estimates in these comparisons, supporting the practical value of its complementary pretraining strategy.

References

[1] K. Park, Y. Na, Y. R. Choi, H. Ryu, H.-W. Shin, and H.-S. Kim. SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structures. ICML, PMLR 306:96917–96942, 2026.

AuthorsOpenReview ↗

Response to Reviewer 2diD (Part 2)

W2 / Q1: Practical value of pretrained model composition

W2: Given that SleepMaMi learns both private and shared representations, and existing sleep foundation models already learn 20k (SleepMaMi), 65k (SleepFM), and 235k (SleepFM-2) PSG cases, it is unclear if composing separately pretrained modality-specific encoders benefits in practice. Is it practical to assume that EEG or ECG foundation models separately trained without using these large PSG datasets add significant value?

Q1: What if SleepXA adopts SleepMaMi’s architecture and learning methods to integrate the pretrained EEG and ECG models?

Response.

Thank you for raising these related questions. Our motivation is to reuse physiological knowledge learned from diverse unimodal recordings. We assess its practical value and the role of the integration method through three complementary comparisons.

(1) Pretrained initialization adds value. We vary only EEG/ECG encoder initialization, keeping the architecture, PSG alignment protocol, and downstream evaluation fixed (Tables R3–R4; manuscript Tables A.13–A.14). Initializing both encoders improves WSC staging MF1 over Scratch by 3.09/5.77 percentage points under LP/FT and substantially improves AF recognition, outperforming either single-encoder initialization.

Table R3. WSC sleep staging. Acc. and MF1 are percentages.

ModeInitializationAcc.MF1
LPScratch81.4073.00
EEG-FM only82.3673.79
ECG-FM only80.7772.49
EEG+ECG-FM (SleepXA)85.1676.09
FTScratch79.9371.84
EEG-FM only83.6776.04
ECG-FM only81.1971.83
EEG+ECG-FM (SleepXA)85.9577.61

Table R4. CinC2017 AF recognition (AF-F1, %).

InitializationLPFT
Scratch18.1039.60
EEG-FM only12.9041.50
ECG-FM only39.8080.80
EEG+ECG-FM (SleepXA)74.3082.30

(2) Independent prediction averaging gives task-dependent gains. We adapt LaBraM and CSFM independently using the same downstream splits and average their class probabilities equally. LP freezes each encoder; FT updates each encoder and head independently. Averaging reduces staging MF1 relative to LaBraM under both protocols. Apnea AUPRC improves under LP but slightly decreases under FT (Table R5). These task-dependent gains motivate joint adaptation on synchronized PSG.

Table R5. WSC prediction averaging without joint PSG alignment.

Task / metricModeLaBraMCSFMAverage
Staging MF1 (%)LP60.430.860.1
FT70.023.767.7
Apnea AUPRCLP0.3140.2460.372
FT0.3880.3230.384

(3) SleepMaMi-inspired integration is feasible, but the integration method matters. To address Q1, we convert LaBraM/CSFM outputs into token sequences and adopt a shared–private design with two-layer modality-private Transformer branches. This adapts SleepMaMi's integration design to pretrained representations; it does not reproduce its full raw-signal micro-/macro-pretraining pipeline. We also include the original SleepMaMi baseline (Table R6).

SleepXA outperforms the original SleepMaMi baseline on all four metrics under LP and FT. The inspired variant matches SleepXA's LP AF-F1 (74.30%) but yields lower WSC staging, AHI, and OSA scores under both protocols. Its FT AF-F1 is 81.30%, versus 82.30% for SleepXA.

Table R6. WSC staging/AHI/OSA and CinC2017 AF recognition. F1 scores are percentages.

ModeMethodWSC staging MF1WSC AHI R²WSC OSA MF1AF-F1
LPSleepMaMi71.510.83970.5165.05
SleepMaMi-inspired67.700.77166.0474.30
SleepXA76.090.86475.0574.30
FTSleepMaMi77.090.86874.9572.68
SleepMaMi-inspired69.820.76467.9581.30
SleepXA77.610.87878.4682.30

Together, these results support the practical value of pretrained encoder reuse and the importance of how their representations are integrated.

AuthorsOpenReview ↗

Response to Reviewer 2diD (Part 3)

W3: Annotation requirements and alternative stage guidance

W3: Dependence on sleep-stage labels raises concerns about scalability and reliability, including annotation cost and inter-scorer variability.

Response.

Thanks for raising this. To investigate the role of “annotation” during alignment, We compare four alignment settings: (i) Full (SleepXA) uses expert labels from SHHS, MESA, and CFS; (ii) None retains instance contrastive learning, reconstruction, and other objectives that do not require sleep-stage labels; (iii) Hybrid retains expert SHHS labels and uses pseudo-labels for MESA/CFS generated by a model trained only on SHHS; and (iv) Partial retains expert SHHS labels while treating MESA/CFS as unlabeled. The only SHHS SleepXA achieves 83.02% pooled staging accuracy on the MESA/CFS validation sets. In the partial setting, unlabeled recordings still contribute to objectives that do not require stage labels.

Removing stage guidance (None) modestly reduces WSC sleep-staging performance. Hybrid and partial guidance both recover part of the LP performance gap (Table R7)

Table R7. Effect of sleep-stage guidance on WSC sleep staging.

ModeGuidanceAcc. (%)MF1 (%)κ
LPNone83.7174.920.751
LPHybrid84.6975.890.762
LPPartial84.2375.570.760
LPFull (SleepXA)85.1676.090.765
FTNone85.4076.350.770
FTHybrid85.5177.190.774
FTPartial84.9076.200.765
FTFull (SleepXA)85.9577.610.780

For whole-night AHI estimation and OSA severity classification, both alternatives improve MAE and classification accuracy over no stage guidance under LP and FT, although full-label guidance remains best (Table R8).

Table R8. Effect of sleep-stage guidance on WSC AHI estimation and OSA severity classification. Acc. and MF1 evaluate OSA severity classification.

ModeGuidanceMAERMSER²Acc. (%)MF1 (%)
LPNone3.6556.5840.85373.4173.68
LPHybrid3.4896.3560.86378.3173.88
LPPartial3.5086.5560.85477.5173.73
LPFull (SleepXA)3.4416.3410.86479.1275.05
FTNone3.6216.6110.85278.7275.19
FTHybrid3.2806.4920.85780.7277.83
FTPartial3.4106.5080.85680.3177.31
FTFull (SleepXA)3.1206.0080.87882.7378.46

For AF recognition, removing stage guidance causes a larger LP performance drop. Hybrid and partial guidance recover most of this gap, while differences across settings narrow after FT (Table R9).

Table R9. Effect of sleep-stage guidance on CinC2017 AF recognition.

ModeGuidanceChallenge-F1AF-F1AF-AUROC
LPNone59.5160.2487.40
LPHybrid72.9173.3589.60
LPPartial72.6072.1288.60
LPFull (SleepXA)73.1874.3091.15
FTNone82.6981.8598.30
FTHybrid82.7181.5698.20
FTPartial82.3981.4198.10
FTFull (SleepXA)83.3582.3098.40

These results demonstrate that although full sleep-stage guidance achieves the best performance, alternative solutions such as pseudo-label and partial-label guidance can achieve performance comparable to the full version.

Summary

Through the added SleepMaMi comparisons, the SleepMaMi-inspired integration study, and the EEG–ECG integration experiments, we aim to further clarify the practical value of pretrained model composition in SleepXA. We also clarify its requirements for sleep-stage guidance and the alternatives available when expert annotations are limited. We hope this additional evidence supports a more favorable reassessment of our work.

We will incorporate these additional comparisons and analyses into the revised manuscript. Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Authors' Responses

Dear Area Chair and Reviewers,

We sincerely thank all three reviewers for their constructive feedback, and the Area Chair for their time and consideration. Especially we sincerely thank all three reviewers for their constructive feedback and for recognising the clarity and quality of presentation (ArTT, KjKF), breadth and depth of experimental evaluation (all three reviewers), and novelty and potential of the compositional foundation-model paradigm (2diD, ArTT).

To facilitate the final assessment, we briefly summarise the key concerns and the additional evidence provided during rebuttal. Revisions to the manuscript are highlighted in blue.

Key Concerns and Our Responses

  • Sleep-stage supervision and scalability (all three reviewers). We appreciate the reviewers’ concerns about the need for sleep-stage labels during pretraining, particularly given the difficulty of obtaining expert annotations at scale. Following Reviewer 2diD’s suggestion (with many thanks), we conducted additional experiments with no, partial, and pseudo-label guidance (Section 4.3; Appendix A.4.9). SleepXA remains trainable without sleep-stage labels. and partial and pseudo-label guidance preserve much of the performance, compared to full supervision. For example, WSC sleep-staging Macro-F1 under linear probing reaches 75.89% with pseudo-labels, compared with 76.09% under full supervision. These results provide practical alternatives when comprehensive expert annotations are difficult to obtain.
  • Novelty and practical value of pretrained model composition (2diD, ArTT). We added quantitative comparisons with SleepMaMi, evaluated SleepMaMi-inspired integration, and conducted controlled experiments on pretrained initialization and alternative fusion strategies. On WSC sleep staging, SleepXA achieves 76.09% Macro-F1 under linear probing, compared with 71.51% for SleepMaMi. These experiments help distinguish the value of reusing independently pretrained physiological knowledge from that of the composition strategy itself.
  • Validation of common--specific learning (ArTT, KjKF). We clarified the motivation for individual objectives, expanded the ablation analyses, and added direct representation-level probes (Appendix A.6). The common branch better supports sleep-stage decoding, whereas the specific branch better preserves modality identity and the examined physiological attributes. These results provide preliminary evidence consistent with the intended decomposition.
  • Evaluation reliability, efficiency, and limitations (ArTT, KjKF). We clarified differences between our SleepFM evaluation and its original published protocol, reported model parameters, inference memory, and alignment-training costs, and added multiseed evaluations for daytime transfer tasks. We also revised the interpretation of clinical prediction results to acknowledge limitations associated with smaller cohorts.

Overall, SleepXA explores an alternative to training multimodal sleep foundation models from scratch: composing independently pretrained physiological models through synchronized PSG. Across 27 downstream tasks, together with the new controlled experiments, our results support the practical value of this compositional approach while clarifying its supervision requirements and limitations.

We hope these additional experiments and clarifications will be helpful to the Area Chair in the final assessment.

Thank you again for your time and consideration.

Best regards,

The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.