List of Papers Browse by Subject Areas Author List
Abstract
Recent breakthroughs in deep learning have catalyzed the evolution of multi-label Electrocardiogram (ECG) classification for cardiovascular diseases. However, prevailing paradigms often suffer from feature fragmentation, which limits their efficacy in identifying complex comorbidities. Specifically, it manifests as a discontinuity in representation, driven by the scale discrepancy between local morphological details and global heartbeat rhythms. Addressing the difficulty of aggregating these fragmented representations for diagnosing complex comorbidities, we propose the Multi-scale Spectral-Semantic Network (MSS-Net), a framework comprising three key components: Adaptive Gated Spectral Injection (AGSI) module, Dual Cross-scale Interaction (DCI) module, and Prior-Guided Feature Aggregation (PGFA) module. Specifically, the AGSI module modulates rebalanced global features in the frequency domain to capture long-range rhythmic dependencies, while DCI module establishes dual interaction channels to synergize local morphological details with global rhythmic contexts, effectively mitigating feature fragmentation to capture features across scales. Subsequently, the PGFA module incorporates a co-learning mechanism that uses individualized clinical contexts as auxiliary supervision signals during training to calibrate complex pathological features. Extensive experiments on three datasets show the effectiveness of our method. MSS-Net outperforms the state-of-the-art by +2.03\% in AUC, indicating robust discriminative capability in complex clinical scenarios.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4140_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/origami-one/MSS-Net
Link to the Dataset(s)
PTB-XL: https://physionet.org/content/ptb-xl/
CPSC2018: http://2018.icbeb.org/Challenge.html
Shaoxing: https://figshare.com/collections/ChapmanECG/4560497
BibTex
@InProceedings{CheYul_MSSNet_MICCAI2026,
author = { Chen, Yuling AND Xia, Jiahao AND Yao, Xingwang AND Luo, Zhuangye AND Zhao, Tianli AND Zeng, Feng},
title = { { MSS-Net: Learning Hierarchical Multi-scale and Multimodal Representations for Complex Cardiac Disease Classification } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16885},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a framework named MSS-Net to address the “feature fragmentation” problem in multi-label electrocardiogram (ECG) classification. The framework integrates three modules: (1) Adaptive Gated Spectral Injection (AGSI), which aims to capture long-range heart rhythm dependencies through frequency-domain modulation; (2) Dual-scale Cross Interaction (DCI), for fusing local morphological details and global heart rhythm context; (3) Prior-guided Feature Aggregation (PGFA), which utilizes medical prior knowledge from clinical text descriptions to calibrate model predictions.
- 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.
- Clear motivation: The paper’s exposition of the “feature fragmentation” problem is intuitive and reasonable, and it effectively illustrates how this issue affects the diagnosis of complex comorbidities (such as MI) through visualization (Figures 1 and 4).
- Robust experimental results: Experiments on three large datasets, PTB-XL, CPSC2018, and Shaoxing, demonstrate that MSS-Net outperforms the compared baseline methods in both AUC and F1 score, with significant performance improvements.
- Comprehensive ablation study: The ablation experiments (Table 2) systematically verify the effectiveness of the three modules, AGSI, DCI, and PGFA, respectively, and prove their contributions to the final 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 core issue of this paper is its lack of innovation, resembling a “patchwork model” rather than original work. The three proposed modules are essentially direct applications or simple combinations of existing mature technologies, lacking in-depth innovation.
- 1.1 The AGSI module: Utilizing Fourier transform for frequency domain analysis to capture global dependencies is a standard practice in signal processing. In deep learning, combining frequency domain information with spatial/temporal domain information is also not a new concept. The design of AGSI is more like grafting a ready-made frequency domain filter onto the backbone network, and its “adaptive gating” design is rather conventional, failing to demonstrate unique advantages over existing frequency domain modeling methods.
- 1.2 The DCI module: Multi-scale feature fusion is a classic paradigm in computer vision and signal processing. Using attention mechanisms for cross-scale feature interaction is also common. Fan et al. explicitly proposed a multi-scale CNN fusion strategy for ECG analysis in their 2018 IEEE JBHI paper titled “Multiscaled fusion of deep convolutional neural networks for screening atrial fibrillation from single lead short ECG recordings”. The DCI module does not introduce a new fusion mechanism; its so-called “dual-channel” design is essentially similar to many existing feature pyramid networks (FPN) or U-Net style skip connections, merely expressed in a different way.
- 1.3 The PGFA module: Utilizing external text information (such as reports) as a multimodal supervision signal to enhance ECG classification has been clearly proposed and deeply explored in recent work. Chen et al. published “Deri: Cross-modal ecg representation learning with deep ecg-report interaction” in IJCAI 2025, constructing a deep ECG-report interaction framework. In contrast, the PGFA module of MSS-Net appears overly simplistic: it merely uses a frozen ClinicalBERT to extract fixed category-level priors and fuses them through a one-time cross-attention. This “plug-and-play” approach lacks in-depth modeling of the complex alignment relationship between the two modalities, resembling more of a borrowing of the capabilities of existing powerful language models (LLMs) rather than proposing a novel multimodal learning paradigm. 2.Lack of in-depth comparison with the most relevant works: Although the Related Work section of the paper mentions some multi-scale and multi-modal works, it fails to conduct a quantitative or qualitative in-depth comparison with directly related and influential works such as Fan et al. (2018) and Chen et al. (IJCAI 2025). This makes it difficult for readers to determine where the true progress of MSS-Net lies in comparison to these specific methods. 3.Weak technical depth and theoretical support: The paper mostly stays at the engineering implementation level and lacks a deep exploration of the underlying mechanisms of why the proposed method works. For instance, why does frequency-domain modulation perform better than sequence models like Transformer in addressing long-range dependencies? Why can simple cross-attention effectively integrate text priors? These questions remain unanswered convincingly. 4.Doubts about computational efficiency and practicality: The paper introduces frequency-domain transformation, additional attention modules, and a large language model (even if it is frozen), which undoubtedly increases the computational cost and deployment complexity of the model. However, there is no discussion at all about the model’s parameter count, inference speed, or memory usage, which is a significant omission for a work claiming to be oriented towards clinical applications.
- 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 provide sufficient information for 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?
The description of the implementation details of the method is inadequate, raising concerns about the reproducibility of the method itself.
- 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?
1.The innovation of the method is insufficient. 2.The description of the implementation details of the method is inadequate, raising concerns about the reproducibility of the method itself. 3.The technical depth and theoretical support are weak. 4.The computational efficiency and practicality are questionable.
- Reviewer confidence
Very confident (4)
- [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 introduces a Multi-scale Spectral-Semantic Network designed to address feature fragmentation in complex cardiac disease classification. The framework synergizes local morphological details with global rhythmic contexts by integrating three novel modules: AGSI for frequency-domain modulation, DCI for dual cross-scale interaction, and PGFA for calibrating representations with medical priors from ClinicalBERT. Extensive experiments on the PTB-XL, CPSC2018, and Shaoxing datasets demonstrate that MSS-Net significantly outperforms current state-of-the-art methods, achieving a 2.03% AUC improvement on PTB-XL.
- 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.
1.The model effectively mimics the diagnostic process by integrating morphological details with rhythmic contexts and medical priors, providing a more holistic approach to complex cardiac disease classification.
2.Extensive evaluation across three diverse public datasets demonstrates consistent performance gains and the model’s superior discriminative capability in real-world clinical scenarios.
- 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 motivation of “feature fragmentation” remains conceptually ambiguous as the authors fail to provide a formal mathematical definition or a quantitative metric to verify its existence in baseline models.
2.Utilizing the Fourier Transform in the AGSI module to capture rhythmic dependencies is theoretically questionable, as global frequency modulation often fails to account for the highly non-stationary nature of pathological ECG signals.
3.The DCI module introduces significant architectural complexity through cross-scale attention and adaptive residual injection, yet the paper lacks a detailed FLOPs or latency analysis to justify this overhead against marginal performance gains.
4.The inclusion of class-level clinical descriptions in the PGFA module during training is problematic, as these textual “priors” likely act as a proxy for the ground-truth labels, potentially leading to significant data leakage.
- 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 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.
(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?
While the performance gains are notable, the paper suffers from conceptual looseness. The central motivation feature fragmentation is never rigorously defined or measured. Furthermore, the reliance on Global Spectral Injection for rhythmic modeling is theoretically questionable given the non-stationary nature of ECG signals. Finally, the potential data leakage in the PGFA module and the lack of a computational complexity analysis make it difficult to assess the true algorithmic contribution of MSS-Net beyond heavy-parameter ensemble learning.
- Reviewer confidence
Very confident (4)
- [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 #3
- Please describe the contribution of the paper
The authors propose MSS-Net, a deep learning architecture for multi-label ECG classification. The framework consists of three main components designed to handle feature fragmentation: 1) the Adaptive Gated Spectral Injection (AGSI) module, which modulates global features in the frequency domain; 2) the Dual Cross-scale Interaction (DCI) module, which uses cross-attention to fuse rhythmic context with morphological details; and 3) the Prior-Guided Feature Aggregation (PGFA) module, which uses class-level textual descriptions encoded by ClinicalBERT to calibrate the final predictions.
- 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 model is evaluated on three standard public datasets (PTB-XL, CPSC2018, Shaoxing) and demonstrates consistent improvements over recent baselines (e.g., MTA-Net, ECG-Mamba). The AUC and F1 gains are solid. Table 2 provides a clear, module-by-module breakdown of performance gains. The ablation isolates the impact of AGSI, DCI, and PGFA, showing that the cross-scale interaction (DCI) drives the largest performance leap, which aligns well with the paper’s claims about bridging semantic gaps. Integrating frequency-domain statistics (AGSI) with time-domain morphological features is a highly practical and clinically sound approach for ECG signals, given their inherent periodic nature.
- 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 architecture reads somewhat like an assembly of existing techniques (Fourier transforms, cross-attention, GRN, and BERT embeddings). While the combination is effective for this specific task, the foundational methodological innovation is limited.
- 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 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.
(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?
I recommend a Weak Accept. The paper addresses a highly relevant clinical problem (complex ECG classification) using a heavily engineered but empirically successful framework. The integration of spectral analysis, cross-scale attention, and text-based priors is logically sound and achieves strong quantitative results across three benchmark datasets. While the methodological novelty leans toward the incremental combination of existing modules, the rigorous experimental validation justifies acceptance.
- 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.
I’m not an expert in this field. Most of my research is in the area of medical images. Therefore, I can’t guarantee 100% accuracy. Initially, I thought this work was a combination approach, but after the author’s rebuttal and based on their explanation, this research is a solution to a specific problem. So I think it’s acceptable, but I’m not confident.
Author Feedback
We sincerely thank the reviewers and the AC for their positive evaluation and valuable suggestions. We appreciate the recognition of MSS-Net as having a “clear motivation” and “extensive and comprehensive evaluation,” showing “robust and effective results.” We summarize our clarification into four key points.
Q1.Definition of Feature Fragmentation and Design Innovation We clarify that feature fragmentation refers to the semantic inconsistency and insufficient interaction between local morphological features and global rhythmic features. Although existing methods perform feature fusion, simple concatenation is often inadequate for modeling interactions between heterogeneous features. For instance, linearly combining isolated features such as irregular R-R intervals, absent P-waves, and a widened QRS may result in a misdiagnosis of concurrent AFib and PVCs. However, in reality, the rhythmic context redefines the morphological semantics. If that widened QRS strictly follows an irregular long-short R-R sequence, the true diagnosis is AFib with Ashman Phenomenon. To resolve this, MSS-Net translates this clinical reasoning into a progressive three-stage computational pathway. First, AGSI modulates deep latent features to provide global rhythmic context. Next, DCI models morphology-rhythm interaction by treating morphology-aware features and rhythmic context as KV and Q, respectively, with learnable weights to retrieve local details under global rhythm guidance. Finally, PGFA further uses fixed textual priors to guide signal representations toward disease-level semantics. Extensive ablation studies confirm that removing any stage disrupts this progressive chain, severely degrading performance. This proves MSS-Net is not an arbitrary module stack, but a synergistic design targeted at resolving feature fragmentation. Regarding Fan [5] and Chen [3], Fan et al. focus on single-lead AF screening and fuse multi-scale convolutional features, which differs from our AGSI modulation and DCI-based morphology-rhythm interaction for multi-class 12-lead diagnosis. Chen et al. use sample-level reports for ECG understanding, whereas PGFA uses class-level disease descriptions and does not rely on patient-specific reports. These distinctions confirm that MSS-Net is not a trivial assembly of borrowed methods, but a novel, targeted architecture.
Q2.Data Leakage The textual data used in our experiments do not contain diagnostic conclusions or disease-category labels. They only provide disease-related physiological and diagnostic feature descriptions. These descriptions are class-level rather than sample-level and therefore do not introduce sample-specific label leakage.
Q3.Complexity and Efficiency While efficiency is valuable for clinical deployment, diagnostic accuracy remains the primary bottleneck in complex ECG comorbidity diagnosis. Following top-tier methodologies (e.g., Chen [3], MTA-Net[16]), standard protocols prioritize discriminative metrics (AUC/F1) over strict complexity analysis. Nevertheless, as stated in the Implementation Details, on a single RTX 4090, MSS-Net requires 16.1 ms per sample for inference and consumes 2.9 GB memory. With ClinicalBERT frozen as a fixed prior, MSS-Net contains only 3.56M trainable parameters. Being lighter than efficient baselines like ECG-Mamba[18] (4.05M) confirms our framework is a highly efficient architecture.
Q4.Non-stationary Nature of ECG We agree that pathological ECG signals are non-stationary. However, AGSI is not a standalone raw-signal frequency classifier. It operates on latent ECG features after the morphological encoder and learns to provide complementary rhythm-aware modulation. The residual connection preserves the original time-domain representation, so local non-stationary morphology is retained.
We hope our responses address the reviewers’ concerns. We will release the code and include additional complexity comparisons in the repository to facilitate reproducibility.
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 multi-module fusion framework (MSS-Net) for ECG classification, integrating frequency-domain modeling, cross-scale feature interaction, and textual priors to alleviate the feature fragmentation problem. The method achieves consistent performance improvements across multiple datasets, with a relatively thorough experimental design and clear ablation studies, demonstrating practical value. However, the novelty is limited, as the approach largely combines existing techniques, and the motivation and design of several modules require further clarification to be fully convincing.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
The rebuttal provides useful clarifications regarding the intended motivation, the use of class-level textual descriptions, and the reported computational efficiency. In particular, the clarification that the textual priors are class-level rather than sample-specific substantially alleviates the concern about direct data leakage.
However, the reviewers’ main concerns remain unresolved based on the submitted manuscript. 1.The methodological novelty appears limited. The three modules are built primarily from established techniques, including Fourier-domain modulation, cross-scale attention, and frozen language-model embeddings. The rebuttal explains their intended progressive interaction, but does not establish that this combination constitutes a sufficiently distinct methodological contribution beyond prior multi-scale and multimodal ECG approaches. 2.“feature fragmentation” remains a qualitative motivation. The rebuttal provides an intuitive clinical example, but does not identify a formal definition, quantitative characterization, or existing analysis in the manuscript that demonstrates this phenomenon in baseline models. 3.The rebuttal distinguishes MSS-Net from Fan et al. and Chen et al. mainly through differences in task setting and the granularity of textual information. These distinctions are relevant, but the submitted manuscript still lacks an in-depth comparison that makes the specific technical progress over the most closely related works clear. 4.The efficiency clarification is helpful, but the manuscript does not sufficiently analyze the computational overhead and deployment implications of the proposed components. Any promised future code release or additional complexity comparison cannot be considered in the present decision.
Overall, the empirical results are promising, and some concerns are clarified by the rebuttal. Nevertheless, the central concerns regarding methodological novelty, technical depth, and support for the core motivation remain insufficiently addressed in the submitted work. I therefore recommend rejection.
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 author’s response addresses the reviewers’ concerns.
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.
This paper proposes MSS-Net for multi-label ECG classification by combining spectral modulation, cross-scale interaction, and class-level textual priors. The manuscript reports competitive performance on three public datasets, with clear gains on several key metrics, and the ablation study supports the contribution of the proposed modules.
I agree with the reviewers that the methodological novelty is moderate and that the feature fragmentation is not directly quantified. Nevertheless, the rebuttal sufficiently clarifies the main concern regarding potential label leakage: the textual priors are class-level descriptions rather than patient-specific reports or ground-truth labels. Overall, the paper addresses a relevant problem and provides adequate empirical support. I therefore consider it marginally above the acceptance threshold.