List of Papers Browse by Subject Areas Author List
Abstract
Sleep physiology arises from the coordinated dynamics of the central nervous system (CNS) and autonomic nervous system (ANS), as reflected by multimodal polysomnography signals including EEG, EOG, EMG, ECG, and respiration. However, existing sleep foundation models often fuse heterogeneous biosignals in a topology-agnostic manner, overlooking their physiological organization. We introduce Omni-Sleep, a sleep foundation model that uses the CNS/ANS partition as a physiological prior for topology-constrained representation learning. Omni-Sleep learns structured representations through three objectives: intra-system consistency, which captures shared subsystem-level factors within neural and cardio-respiratory signals; inter-system synchronization, which aligns subsystem trajectories to model brain–body dynamics; and latent-space masked temporal modeling, which captures long-horizon sleep dynamics. Pre-trained on over 100,000 hours of multi-center multimodal PSG data, Omni-Sleep is evaluated on sleep staging and multi-disease classification. Across datasets and modality-ablation settings, Omni-Sleep outperforms strong foundation-model baselines, showing improved label efficiency, cross-dataset generalization, and robustness to missing modalities. These results highlight the value of physiological hierarchy for generalizable sleep representation learning. Code is available at \url{https://github.com/AutoBrain-sleep/OmniSleep}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2428_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AutoBrain-sleep/OmniSleep
Link to the Dataset(s)
N/A
BibTex
@InProceedings{HouZho_OmniSleep_MICCAI2026,
author = { Hou, Zhoujie AND Wang, Song AND Lou, Kexin AND Wang, Mo AND Wei, Chen AND Liu, Quanying},
title = { { Omni-Sleep: A Sleep Foundation Model via Hierarchical Contrastive Learning of CNS–ANS Dynamics } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16894},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a foundational model with a hierarchical pretraining framework that explicitly separates central nervous system (CNS) and autonomic nervous system (ANS) representations to enforce both intra-system consistency and inter-system coupling. The framework further incorporates a long-horizon latent mask prediction task to promote temporally consistent representations over extended contexts. Experimental results demonstrate strong performance in out-of-domain generalization, few-shot sleep stage transfer learning, and multi-disease risk stratification across multiple datasets.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
- Introduces a novel pretraining strategy for sleep foundation models by explicitly separating central nervous system (CNS) and autonomic nervous system (ANS) signals, which is conceptually well-motivated.
- Demonstrates improvements over state-of-the-art methods under linear probing and few-shot fine-tuning settings on unseen datasets, suggesting strong representation learning capability.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
- The evaluation of out-of-domain generalization and few-shot fine-tuning is limited to a subset of datasets. It is unclear why these experiments were not conducted across all available out-of-domain datasets, raising concerns about whether the conclusions would remain consistent.
- The description of the model architecture lacks clarity and sufficient detail. In particular, the MAE blocks, although shown in Figure 2, are not described and justified in the manuscript.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a conceptually interesting and well-motivated approach, particularly through the explicit separation of CNS and ANS representations within a hierarchical pretraining framework. This design choice is novel in the context of sleep modeling and has the potential to provide a meaningful inductive bias. The incorporation of long-horizon latent mask prediction is also a sensible addition to encourage temporal consistency in the learned representations.
Empirically, the method demonstrates promising results, including improvements over prior work under linear probing and few-shot transfer settings, as well as encouraging performance in out-of-domain generalization. Although the number of out-of-domain datasets is limited, the results suggest that the learned representations are robust and transferable across datasets and tasks.
However, one key concern limits the strength of the contribution. The experimental evaluation is not sufficiently comprehensive, particularly for out-of-domain generalization and few-shot settings, which are demonstrated on only a subset of datasets rather than across all available datasets.
Overall, while the paper introduces a promising idea and shows encouraging results, the current presentation and evaluation leave some uncertainty.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The authors have addressed most of my concerns except addressing my suspicions on the out of domain experiment results.
Review #2
- Please describe the contribution of the paper
The paper proposes Omni-Sleep, a multimodal sleep foundation model that separates PSG signals into CNS and ANS branches and pretrains them with a hierarchical objective combining intra-system contrastive learning, inter-system alignment, and long-horizon masked prediction. The method is evaluated on sleep staging and sleep-related disorder prediction across multiple datasets and modality subsets, with the goal of improving robustness under label scarcity, domain shift, and missing modalities. This framing is consistent with the paper’s stated contributions and experimental setup.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper addresses a relevant real-world problem in multimodal sleep modeling, namely how to leverage heterogeneous PSG signals under missing modalities and limited labels. A clear strength is the broad empirical scope: the authors pretrain on a large multi-center corpus and evaluate on both out-of-domain sleep staging and disease risk prediction tasks. The reported results are consistently stronger than the included foundation-model baselines across EEG-only, CNS, ANS, and full-modality settings, and the paper also includes a few-shot evaluation and one architectural ablation. Overall, the work is ambitious, practically motivated, and supported by a fairly extensive experimental section.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The central “physiology-aware / topology-aware” claim feels stronger than the evidence provided. In practice, the method appears to use physiology mainly as a manual grouping prior: modalities are split into CNS (EEG/EOG/EMG) and ANS (ECG/respiration), trained with separate objectives, and then aligned. The paper does not provide evidence that the learned representations capture physiologically meaningful structure beyond this hand-designed split. In addition, the intra-system contrastive loss defines each modality’s positive target as the mean representation of the other modalities in the same subsystem, which pushes embeddings toward shared information within the group and may reduce complementary modality-specific information. As written, the method reads more as a structured multimodal training scheme than as a genuinely physiology-informed model. 2.A key methodological component the long-horizon stage is under-specified. The macro-scale masked prediction objective is presented as one of the main novelties and is illustrated with 1.5-hour windows, but the paper does not explain why this duration was chosen, whether windows overlap, what stride is used, how masking is sampled, or whether performance is sensitive to these design choices. Since this component is also used later to support claims about clinically meaningful long-range dynamics, these details are important for both interpretation and reproducibility. 3.The disease-prediction protocol is not described concretely enough. The paper states that a held-out subset of 1,159 SHHS subjects is used for downstream disease-risk classification and that binary labels are defined according to established clinical criteria, but the exact operational definitions are not provided. More importantly, it is unclear how these subject-level disease labels relate to the long-horizon windows introduced by the method: whether multiple 1.5-hour windows are extracted per subject, whether each window inherits the same subject label, and how window-level representations or predictions are aggregated into the reported subject-level AUROC. This missing step makes the disease-risk results harder to interpret. 4.The robustness-to-missing-modalities claim is not backed by a sufficiently explicit protocol description. The paper repeatedly emphasizes robustness to sensor dropouts, modality ablation, and partial observability, and reports results for EEG, CNS, ANS, and Full settings. However, it is not clearly explained how inference is performed when only a subset of modalities is available, or which modalities were actually present and usable in each dataset and evaluation setting. Since robustness under missing modalities is one of the paper’s main practical claims, this protocol should be stated much more explicitly. 5.The baseline comparison needs to be specified more clearly to support the main empirical claim. The paper claims consistent improvement over prior baselines, but it is unclear how these baseline results were obtained. For sleep staging, Table 1 compares mainly against SleepFM and SleepGPT across EEG/CNS/ANS/Full settings. For disease prediction, Table 2 compares against SleepGPT and a demographic baseline. However, the paper does not clearly state whether the prior foundation models were evaluated from released checkpoints or retrained by the authors. If retrained, whether they used the same data access, modality subsets, and downstream probing/fine-tuning protocols; or which modality subsets each baseline can actually consume. In addition, only Omni-Sleep’s own size is reported (56M parameters), while model size or compute for the compared baselines is not given. Finally, in the EEG-only setting, the comparison would be stronger if it also included strong task-specific supervised baselines like EEGNet/EEGConforfer/etc., rather than only other foundation models. 6.Practical reproducibility is limited in the current form. Although an anonymized code link is provided in the abstract, the released repository is not fully self-contained in its current state and appears to require additional files/checkpoints to run the provided scripts. Together with the relatively sparse implementation detail in the manuscript, this makes the work difficult to reproduce.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper presents a relevant idea and fairly broad experiments, and the reported numbers against the included baselines are promising. However, I remain unconvinced by the central physiology-aware/topology-aware claim, because the method mainly implements a coarse CNS/ANS grouping and does not demonstrate that it learns genuinely physiological structure beyond that architectural split. In addition, key experimental details are missing, reproducibility appears incomplete, ablations are too limited for a method paper, and the downstream evaluation lacks stronger task-specific baselines in at least the EEG-only classification setting. These issues make the paper interesting but not fully convincing in its current form.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal sufficiently addresses my main technical concerns to move the paper above the acceptance threshold. In particular, the authors clarify that the CNS/ANS split is a hand-designed physiological prior rather than a discovered topology, provide missing MAE/RoFormer hyperparameters and implementation details, and clarify the disease-prediction protocol with subject-level aggregation and one AUROC prediction per subject. Some concerns remain regarding physiological probing, limited ablations, task-specific baselines, and reproducibility, but I view these as camera-ready / limitation issues rather than reasons for rejection after the rebuttal.
Review #3
- Please describe the contribution of the paper
This paper introduces Omni-Sleep, a novel structure-aware foundation model for sleep analysis. The model captures the interplay between the Central Nervous System and Autonomic Nervous System through a unique hierarchical contrastive learning framework. And the model demonstrates SOTA performance with significant improvements in out-of-domain generalization, label efficiency across diverse sleep-related tasks.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Omni-Sleep adopts a highly novel hierarchical pre-training strategy. It first learns the intra-modal consistency of various physiological signals within the CNS and ANS in short-range temporal windows, then aligns the interactions and dynamic relationships between the two physiological systems across long sequences. And a masked autoencoder(MAE) is utilized to enhance the model’s capability in capturing long-range dependencies. Omni-Sleep follows physiological principles. The network does not use complex theories or network frameworks. Based on contrast loss, it establishes consistency within the system and achieves synchronization across systems, capturing the complex brain-body dynamics during sleep.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1,Key hyperparameters for physiological time series forecasting, such as the sequence length T and L, and feature dimension D, are not specified in the main text. 2,The current ablation experiments are insufficient. Only the removal of the RoFormer module is evaluated, while RoFormer is not the core contribution of this work. Moreover, as a vital component of both the Encoder and MAE, it’s unclear what alternative structure is adopted to replace RoFormer in the ablation setting. 3,The baseline comparisons are limited. For binary classification tasks of certain diseases, the AUROC values are even lower than 0.5, indicating that the selected model performance is inferior to random guessing.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper is well-organized and the anonymous code is adequately provided. Following physiological principles, this work proposes a novel hierarchical training paradigm. With contrastive loss, it effectively explores intra-system consistency and cross-system correlation. The network framework is concise and streamlined, with only 56M parameters. In addition, extensive experimental validations on multiple datasets sufficiently demonstrate the advantages and generalization of the proposed method.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
1.Physiological structure beyond a hand-designed split (R2, AC) . We agree that the CNS/ANS partition should be described as a hand-designed physiological prior rather than a structure automatically discovered by the model. We will revise the wording from “discovering physiological topology” to “prior-guided, topology-constrained representation learning.” In Omni-Sleep, the CNS/ANS split defines the representation topology, while the learned structure comes from the objectives: intra-system covariation, inter-system CNS–ANS coupling, and long-horizon temporal evolution during sleep. The intra-system mean target is used to learn shared subsystem-level factors for robustness to sensor dropout, while modality-specific encoders and projection heads still preserve modality-dependent information before alignment. We will soften the claim and discuss representational probing of sleep-stage, cardiorespiratory, and modality-specific factors as an important future analysis. 2.Hyperparameters and MAE/RoFormer details (R1, R2, R3, AC). We will add the missing hyperparameters and implementation details: the long-horizon sequence length is L=180 epochs, corresponding to 1.5 hours, which roughly matches a typical sleep cycle; the embedding dimension is D=256; the MAE mask ratio is 0.5 with contiguous masked blocks of 10 epochs; and the cross-system pooling window is 60 epochs. We will clarify that the MAE module refers to masked autoencoding in the latent space, rather than raw PSG signal reconstruction. It operates on cross-epoch latent embeddings: masked epoch embeddings are replaced by learned mask tokens, processed by an 8-layer RoFormer encoder, and reconstructed by a 4-layer RoFormer decoder only at masked positions using an L1 loss. In the w/o RoFormer ablation, the RoFormer is removed rather than replaced by another temporal model; the modality encoders, projection heads, and hierarchical contrastive objectives are kept unchanged. This ablation therefore isolates the contribution of long-range temporal sequence modeling. 3.Evaluation protocols (R1, R2, AC). For disease prediction, whole-night recordings are split into 1.5 h windows sharing the same subject-level label. Window representations are averaged into one subject-level representation before prediction, and AUROC is computed with one prediction per subject, preventing subject leakage. For baselines, SleepFM and SleepGPT were evaluated from released checkpoints under matched splits and downstream protocols where supported. We will clarify baseline provenance, supported input modalities, label definitions, probing/fine-tuning protocols, and model sizes: SleepFM 4.4M, SleepGPT 134M, and Omni-Sleep 56M. SleepGPT supports EEG/EOG/EMG but not ANS-only inputs, so it cannot be evaluated in the ANS-only setting. For baseline scope and reproducibility, we will clarify that our main goal is to evaluate the representation quality of a sleep foundation model under frozen probing, few-shot transfer, missing-modality inference, and disease-risk prediction, rather than to exhaustively benchmark task-specific supervised sleep-staging models. EEG-only supervised models such as EEGNet and EEG-Conformer are valuable references for EEG-only staging and will be discussed as complementary task-specific baselines where appropriate. However, they address a different question and are not necessary for evaluating multimodal foundation representations. For disease-specific binary classification, low AUROCs for some baselines indicate limited transfer under the matched probing protocol rather than implementation failure. We agree that stronger task-specific baselines would further enrich the evaluation and will consider them in future work. To support reproducibility, after acceptance and in accordance with anonymity and rebuttal rules, we will release environment setup, preprocessing instructions, split files, checkpoint-loading scripts, and model checkpoints.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
This paper proposes Omni-Sleep, a hierarchical sleep foundation model capturing CNS-ANS interactions via contrastive learning and masked autoencoding. Reviewers acknowledge the conceptual novelty, the physiological motivation, and promising out-of-domain and few-shot results.
Authors should focus on addressing: specifying key hyperparameters for sequence length, feature dimension, and long-horizon masked prediction; clarifying details of the MAE and RoFormer modules and alternative structures in ablations; expanding baseline comparisons, particularly for disease-specific binary classification tasks; and providing more comprehensive evaluation across all out-of-domain datasets to strengthen generalization claims. Rebuttal should also clarify how CNS/ANS representations capture meaningful physiological structure beyond hand-designed splits.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper presents a physiologically motivated sleep foundation model that separates CNS and ANS signals and combines hierarchical contrastive learning with long-horizon masked prediction. Reviewers found the problem important, the design conceptually interesting, and the empirical results strong across sleep staging, few-shot transfer, out-of-domain evaluation, and disease-risk prediction. Initial concerns focused on insufficient architectural detail, limited ablations, unclear disease-prediction protocol, baseline fairness, and the strength of the physiology-aware claim. The rebuttal appears to have addressed the major technical concerns, with two reviewers moving to accept and the third already positive. Remaining issues are mostly suitable for camera-ready clarification.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors have provided sufficiently clarifying responses for this paper to be included in MICCAI.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After rebuttal, all reviewers suggest accepting this paper with one reviewer not submmting the final score. Based on the rebuttal, as well as the comments, this paper should be accepted.
It is suggested that the authors sufficiently consider the comments raised by the reviewers and further improve the paper in the final version.
