List of Papers Browse by Subject Areas Author List
Abstract
Automatic sleep staging is a fundamental task in sleep health assessment and clinical diagnosis, yet it remains challenging due to insufficient exploitation of multimodal complementarity, heterogeneous physiological signals, and complex sleep stage transition patterns. This paper proposes CalibSleep, a dual-stream sleep stage classification framework that jointly models time-domain EEG/EOG signals and their corresponding time–frequency representations. CalibSleep employs modality-specific encoders to capture complementary temporal and spectral features, and introduces a cross-modal calibration module to explicitly align heterogeneous representations and adaptively balance modality contributions. Furthermore, physiological prior knowledge of sleep stage transitions is incorporated through a rule-aware classification strategy, improving prediction stability and clinical consistency without relying on complex long-range temporal modeling. Experiments on an in-house clinical dataset from Shanghai Jiao Tong University Affiliated Sixth People’s Hospital and the public SHHS1 dataset demonstrate that CalibSleep achieves accuracies of 89.6% and 88.3%, respectively, outperforming existing methods. Ablation studies further validate the effectiveness of each proposed component. These results indicate that CalibSleep provides an effective and well-generalizable solution for automatic sleep stage classification. The code is publicly available at https://github.com/Cx3300903/CalibSleep.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/1908_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/Cx3300903/CalibSleep
Link to the Dataset(s)
SHHS1: https://sleepdata.org/datasets/shhs
BibTex
@InProceedings{CheXia_CalibSleep_MICCAI2026,
author = { Chen, Xiang AND Shu, Jia AND Wang, Zheyuan},
title = { { CalibSleep: Cross-Modal Calibration and Rule-Aware Learning for Automatic Sleep Stage Classification } },
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 focuses on automatic sleep staging and proposes the CalibSleep framework, a dual-branch cross-modal calibration fusion model. It employs bidirectional attention alignment and adaptive calibration fusion for heterogeneous features in time and time-frequency domains, combined with rule-aware stage transition regularization. The proposed method achieves superior performance over state-of-the-art approaches on both an in-house dataset and the public SHHS1 dataset, with a design that conforms to sleep physiological rules.
- 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.
This paper proposes the CalibSleep, a dual-branch cross-modal calibration fusion framework, for the automatic sleep staging task. By leveraging bidirectional cross-attention to achieve semantic alignment of heterogeneous features in the time and time-frequency domains, combined with adaptive calibration and AASM rule-aware phase transition regularization, it effectively improves the physiological consistency and classification performance of sleep staging. Extensive experiments on both the in-house dataset and the public SHHS1 dataset demonstrate that the proposed method achieves state-of-the-art (SOTA) performance.
- 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 CMC module proposed in this paper achieves semantic alignment through bidirectional cross-attention and uses an adaptive weight α to balance time-domain and time-frequency-domain features. However, only the overall effectiveness of the CMC module is verified via ablation experiments; the individual contributions of bidirectional attention and adaptive calibration are not decomposed, making it impossible to clarify their independent effects on performance improvement. 2.The comparative experiments in this paper only cover several unimodal and time-frequency multimodal baselines, without including high-performance sleep staging methods based on Transformer and graph neural networks in recent years. 3.The specific design of the AASM transition matrix and the sensitivity of the hyperparameter λtrans are not elaborated, nor is the improvement of regularization on each sleep stage quantitatively analyzed.
- 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?
Core Factors Leading to the Total Score 1.Strong Advantages Supporting Acceptance a. The framework is designed to align heterogeneous features while incorporating AASM rule-aware phase transition regularization, which strictly conforms to the physiological rules of sleep staging. This ensures the model’s decisions are interpretable and clinically meaningful, a key highlight of the work. b.Extensive experiments on both an in-house dataset and the public SHHS1 dataset validate that CalibSleep outperforms existing state-of-the-art methods. The competitive results confirm the framework’s effectiveness and robustness. c.The paper provides sufficient visualization and analysis of model decisions, such as attention weights and feature alignment results, which helps readers understand how the model processes sleep data and enhances the credibility of the method. 2.Minor Deficiencies Leading to “Weak Acceptance” a.The CMC module’s overall effectiveness is verified via ablation, but the individual contributions of bidirectional attention and adaptive calibration are not decomposed. This fails to clarify the specific role of each sub-component, reducing the transparency of the technical design. b. Comparative experiments only include traditional unimodal and time-frequency multimodal baselines, lacking direct comparisons with recent high-performance methods based on Transformer and graph neural networks (GNN). This makes it difficult to fully demonstrate the superiority of CalibSleep in the current research landscape. c.The AASM transition matrix’s specific design and hyperparameter λtrans sensitivity analysis are not elaborated. Additionally, the quantitative improvement of phase transition regularization for each sleep stage is missing, weakening the rigor of the physiological prior integration.
- Reviewer confidence
Somewhat confident (2)
- [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
Review #2
- Please describe the contribution of the paper
This paper proposes CalibSleep as a dual-stream framework for automatic sleep stage classification that jointly models time-domain EEG/EOG signals and their time-frequency representations. The method introduces a Cross-Modal Calibration module to dynamically align and balance heterogeneous features alongside a rule-aware classification strategy that incorporates physiological sleep transition priors to improve prediction stability.
- 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 integration of physiological priors via a rule-aware transition loss provides a biologically meaningful constraint to the classification process without relying on complex long-range sequence modeling. The dual-stream architecture effectively captures both temporal and spectral dynamics making good use of modal complementarity. Furthermore the experimental results on both an in-house clinical dataset and the public SHHS1 dataset demonstrate strong empirical performance and reliable generalization capabilities.
- 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.
While the CalibSleep framework demonstrates solid empirical performance, its architectural novelty is somewhat limited. The model primarily relies on a combination of established neural network components, such as the standard stack of CNN, SENet, and BiGRU for the time-domain encoder, and the existing TimesNet for the time-frequency branch. The cross-modal calibration module also utilizes standard cross-attention and a gating mechanism. Although this combination is effective, it represents more of an engineering integration rather than a fundamental methodological innovation in multi-modal deep learning. Additionally, integrating these multiple modules may introduce extra computational overhead, but the paper lacks a detailed analysis of parameter counts and inference latency which are important for clinical deployment. Finally, including comparisons with more recent transformer-based or graph-based state-of-the-art models would further strengthen the demonstration of the proposed method’s absolute superiority.
- 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- 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?
While CalibSleep achieves strong empirical results, its architectural contribution leans more towards an engineering integration of existing components rather than a fundamental methodological innovation in multi-modal deep learning. Furthermore, the absence of a detailed computational efficiency analysis (e.g., parameter count, inference latency) makes it difficult to assess the model’s practical viability for clinical deployment. Given the somewhat limited novelty and the missing comparisons with more recent transformer-based or graph-based baselines, my initial recommendation is a Weak Reject. I would be willing to raise my score if the authors can provide the necessary computational profiling and stronger baseline comparisons during the rebuttal phase.
- 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.
N/A
Review #3
- Please describe the contribution of the paper
This paper proposes CalibSleep, a dual-stream deep learning framework designed for automatic sleep stage classification using EEG and EOG signals. The main contributions are twofold: First, it introduces a cross-modal calibration module (CMC) that explicitly aligns and dynamically balances features from both the time domain and time-frequency representations. Second, it proposes a novel “rule-aware classification” strategy that integrates AASM physiological guidelines into the loss function. This acts as a soft regularization term to penalize physiologically implausible sleep stage transitions between adjacent epochs, achieving high accuracy (89.6% on an in-house dataset, 88.3% on SHHS1) without relying on complex, heavy long-sequence temporal modeling.
- 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 most striking strength of this paper is the elegant design of the rule-aware transition loss. Instead of employing computationally expensive long-range temporal models (such as large Transformers or deep RNNs) to implicitly learn transition probabilities, the authors ingeniously inject clinical rules directly into the objective function. By calculating the transition matrix of adjacent single-epoch predictions and penalizing unlikely shifts, the model captures essential temporal context and clinical consistency in a highly lightweight and effective manner.
- 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.
Robustness of Rule-Aware Loss Against Label Noise: While theoretically appealing, the practical utility of the rule-aware loss requires further justification, specifically regarding the inherent ambiguity of sleep stage transitions. Sleep transitions are notoriously fuzzy, and even highly experienced clinicians often disagree when annotating these boundary epochs. For instance, as detailed in the dataset paper “ISRUC-Sleep: A comprehensive public dataset for sleep researchers”, inter-rater disagreement between expert scorers can be as high as 20%. Given that the ground-truth labels themselves contain significant noise during transitions, enforcing a rigid physiological prior matrix might over-penalize the model or force it to learn hard boundaries that do not reflect the continuous nature of human sleep. The authors need to discuss whether this explicit signal remains effective in the presence of such high inter-rater variability.
Lack of Statistical Significance Testing: The quantitative results presented in Tables 1 and 2 report absolute metrics (ACC, MF1, Kappa) but completely omit standard deviations or confidence intervals. Because the performance margins between the proposed method and some baselines (or among ablation variants) are relatively narrow, it is impossible to determine whether these improvements are statistically significant or merely the result of favorable random seeds. To convincingly prove the method’s effectiveness, the authors must conduct multiple runs to report standard deviations and perform appropriate statistical significance tests (e.g., t-tests or Wilcoxon signed-rank tests).
- 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.
(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?
My recommendation leans towards Borderline Accept / Weak Accept.The paper presents a logically sound architecture. The introduction of the rule-aware loss is a clever, computationally efficient way to integrate clinical priors into a single-epoch classifier, and the cross-modal calibration is well-executed.However, the empirical evaluation currently lacks the rigor expected for publication due to the missing variance measures and statistical significance tests. Furthermore, the core mechanism (the rule-aware loss) rests on an assumption of clear-cut transitions that fails to account for well-documented inter-rater ambiguity in clinical practice.I am willing to raise my score if the authors can provide the necessary statistical significance analyses (mean $\pm$ std across multiple runs) and offer a nuanced discussion/experiment regarding how their transition regularization handles the inherent label noise during sleep stage shifts in their rebuttal.
- 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
We thank the reviewers and AC for the constructive comments and for recognizing CalibSleep’s design, evaluation, performance, and clinical interpretability. Following MICCAI rebuttal policy, we do not introduce new/additional results, but clarify the submitted evidence and acknowledge limitations where new experiments would be required.
Novelty and positioning. We agree that several low-level encoders use established components. Our contribution is not the isolated use of CNN/BiGRU/TimesNet/cross-attention, but the formulation of sleep staging as calibrated integration of temporal and time-frequency evidence under physiological transition constraints. Unlike simple concatenation or one-way attention, CMC performs bidirectional semantic querying between domains and learns an epoch-adaptive calibration gate alpha to balance temporal and spectral reliability. Fig. 3(a) supports this design: alpha shows meaningful stage-dependent behavior, emphasizing time-domain EOG-related evidence in Wake/REM and spectral evidence in N1-N3. Component contribution and transition prior. Table 2 evaluates the proposed components: compared with dual-stream concatenation (0.836/0.833/0.828 ACC/MF1/Kappa), adding CMC improves performance to 0.872/0.846/0.841, and further adding rule-aware regularization yields 0.896/0.877/0.874.We agree that the manuscript should better disentangle sub-component roles: bidirectional attention handles cross-domain alignment and feature conflict; adaptive calibration performs sample-wise evidence weighting; Lcalib stabilizes feature consistency during training; and Ltrans penalizes physiologically implausible adjacent-epoch predictions. M encodes adjacent-stage rationality according to AASM-guided sleep evolution knowledge. Ltrans is a probability-weighted soft regularizer, not a hard constrained decoder. lambda_trans=0.1 balances this prior with cross-entropy: lambda_trans=0 degenerates to ordinary classification, while an overly large value may over-constrain ambiguous transitions.
Baselines, computation, and deployment. We apologize that the baseline positioning was not sufficiently clear. Table 1 includes representative unimodal, time-frequency, multimodal, and Transformer-related methods, including SFormer/SleepTransformer and MultichannelSleepNet variants, as well as XSleepNet, SleepBoost, CoReSleep, and CrossFusionSleepNet. We acknowledge that additional GNN comparisons and parameter/latency profiling would further strengthen the paper. Qualitatively, the physiological priors are deployment-friendly: the contrastive calibration loss and transition loss are used only during training and introduce no extra inference-time decoder. CalibSleep operates on 30-s EEG/EOG epochs and does not require whole-night self-attention or graph construction.
Statistical rigor and robustness. We agree that reporting mean/std and significance tests would improve empirical rigor, and we will explicitly discuss this limitation. Nevertheless, the submitted evidence is not limited to one narrow setting: CalibSleep shows consistent gains on the in-house No.6 dataset and external SHHS1 dataset, across ACC/MF1/Kappa, ablation variants, and interpretability analyses.
Robustness to label ambiguity. We fully agree that boundary epochs are ambiguous and inter-rater disagreement exists. Our rule-aware loss does not assume deterministic transitions or override expert labels. It acts on the predicted transition probability p_{t-1}p_t^T and softly discourages implausible transitions, while cross-entropy from manual annotations remains the primary supervision. Thus, ambiguous epochs can still be determined by signal evidence. Fig. 3(b,c) suggests that the regularizer mainly suppresses physiologically unlikely confusion, such as N3-REM, rather than blindly smoothing predictions.
We hope these clarifications address the concerns on novelty, component roles, baselines, practicality, statistical rigor, and robustness.
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 a dual-stream framework for automatic sleep stage classification. The reviewers acknowledged the model design, extensive experiments, improved performance, and clinically meaningful interpretability. This paper’s significant would be better highlighted if the authors address several important points raised by the reviewers, such as clarifying the methodological novelty, positioning the work relative to stronger recent baselines (through proper background review), disentangling the contributions of the proposed calibration components, reporting variance and statistical significance, discussing computational cost, deployment practicality, and robustness.
- 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 rebuttal seems to have clarified the intended role of the calibration and rule-aware components and to have preserved a positive overall stance after rebuttal, even though some requests for stronger contextualization and significance analysis remained. Because the core contribution is useful and the main concerns appear manageable rather than fatal, I support acceptance.
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 rebuttal addressed the major reviewer concerns satisfactorily, particularly regarding the positioning of the contribution, the role of the calibration components, the interpretation of the rule-aware transition regularization, and the practical deployment considerations. The authors also clarified the rationale behind the architectural design and acknowledged the limitations regarding additional statistical analysis and computational profiling.
The paper presents a technically sound and clinically meaningful framework for automatic sleep stage classification. The integration of cross-modal calibration with physiologically guided transition regularization is well motivated and supported by strong empirical performance on both in-house and public datasets. The rule-aware transition loss provides an elegant and lightweight mechanism for incorporating clinical sleep transition priors without requiring heavy long-sequence modeling.
I recommend acceptance.
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.
The reviewers found the paper clinically motivated and empirically strong. After rebuttal, the main concerns on novelty, baselines, efficiency, and statistical validation were sufficiently addressed, supporting acceptance.