List of Papers Browse by Subject Areas Author List
Abstract
Electrocardiogram (ECG) mass screening plays a crucial role in the early detection of asymptomatic cardiovascular abnormalities, enabling timely intervention and prevention of sudden cardiac events in seemingly healthy individuals. Nowadays, ECG self-supervised learning (eSSL) is the cutting-edge frontier technique for representation learning on unannotated ECG data. Despite extensive efforts in modeling heartbeat rhythms from ECG signals, the continuous temporal morphology-based dynamics of ECG patterns remain incompletely represented in current modeling framework. In this work, we introduce a novel shape-based eSSL approach to capture the intricate wave deflection characteristics of ECG signals, in which temporal trend shifts are discretized and treated as word tokens in a symbolic sequence. We propose ShapeEFM: a novel transformer-based foundation model pretrained using generative eSSL approach. We evaluated ShapeEFM across EchoNext and CSN datasets, where it demonstrated robust competitiveness and enhanced interpretability against the existing eSSL methods. The code is available at: https://github.com/AQ-MedAI/ShapeEFM
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2861_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/AQ-MedAI/ShapeEFM
Link to the Dataset(s)
N/A
BibTex
@InProceedings{LiuHua_ShapeEFM_MICCAI2026,
author = { Liu, Huan AND Fan, Jiaqi AND Guo, Siyan AND Wang, Junyi AND Bai, Xiaoyu AND Lu, Le},
title = { { ShapeEFM: A Shape-Based Pretrained Foundation Model on ECG Morphology Comprehension } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes ShapeEFM, a self-supervised pretraining framework for ECG morphological understanding. Departing from fixed-length or heartbeat-level tokenization, the method detects key waveform trends via Perceptually Important Points (PIPs), defines local morphological fragments as Cardio-Morphemes, and assembles them into Cardio-Sentences. The model combines temporal encoding, spatiotemporal encoding, Transformer architecture, and vector quantization for discrete semantic modeling, with pretraining completed through masked prediction. Experiments on EchoNext and CSN datasets demonstrate competitive linear probing results, supported by ablation studies and latent space visualizations providing preliminary evidence for method effectiveness and partial interpretability.
- 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 motivation of this work is clear and well-grounded. The paper identifies that existing ECG self-supervised methods stress rhythm or high-level semantics while underrepresenting continuous waveform morphology, making the shape-oriented re-design of representations genuinely heuristic.
2.The technical pipeline is complete, forming a coherent flow from morphological unit construction, tokenizer training, vocabulary learning, to masked pretraining. 3.The approach carries certain interpretability, transforming ECG signals into finer-grained linguistic units with ablation experiments and latent space visualizations offering preliminary support for representation learning effectiveness. - 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 boundary of key innovation remains insufficiently defined. The proposed framework closely resembles existing ECG language modeling methods such as HeartLang. At the high-level framework, the primary change appears to substitute heartbeat-level tokenization with PIP-based turning point segmentation, suggesting novelty lies more in segmentation granularity and representational units rather than the pretraining framework itself. 2.The claim of overcoming discrete segmentation limitations appears overstated. The proposed method still relies on explicit segmentation and discretization, merely employing finer-grained morphological segmentation rather than fixed windows or heartbeat cuts, rather than escaping discrete segmentation per se. 3.The robustness of the segmentation strategy requires thorough validation. As HeartLang noted, prior fine-grained segmentation methods suffer from signal quality sensitivity, and even QRS-Tokenizers struggle with irregular QRS complexes and baseline drift. ShapeEFM’s PIP algorithm operates at finer granularity with extreme dependence on local waveform turning points, rendering it more vulnerable to baseline drift, high-frequency noise, and muscle artifacts. Since subsequent Cardio-Morpheme construction, vocabulary learning, and prediction targets all rest upon segmentation quality, boundary instability may undermine entire pretraining reliability, yet the paper currently provides only basic filtering without direct segmentation robustness or noise sensitivity analysis. 4.The generalization from single-lead segmentation to 12-lead settings lacks sufficient justification. The paper explicitly states that PIP discovery executes only on Lead I, then applies the resulting keypoints across all 12 leads. This embeds a strong assumption that Lead I morphological boundaries adequately represent local morphological variations across all leads. However, different leads reflect identical cardiac electrical activity projected along different spatial directions, with local waveform morphology and turning point positions not necessarily strictly consistent. The paper offers no dedicated analysis of this assumption, nor comparison experiments with per-lead segmentation or multi-lead joint segmentation, leaving this design’s rationality inadequately explained. 5.Experimental evaluation depth should be expanded. Current experiments rely primarily on linear probing, ablation studies, and qualitative visualization. This work lacks cross-dataset generalization analysis, systematic robustness testing, and statistical significance testing, rendering existing experiments more indicative of potential than establishing significant advantages over existing methods. 6.Method description and experimental settings could be clarified to guarantee reproducibility. For instance, pretraining data MIMIC-IV-ECG and downstream data EchoNext, CSN exhibit inconsistent sampling frequencies, while the method employs fixed hyperparameters closely related to the temporal scale, such as PIP quantity ratios and maximum Cardio-Morpheme length. The paper fails to clearly clarify whether different sampling rates underwent unified resampling, or how these hyperparameters maintain comparability across different temporal resolutions. 7.In the method section, some reused symbols suffer from inconsistency issues, increasing comprehension and reproduction costs.
- 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.
(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?
This paper could potentially enhance interpretability by linking signal patterns to semantic units. However, the justification for key design choices is not sufficient.
- 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.
Thanks for the authors’ rebuttal. While the contribution of this work still seems incremental, I am willing to change my recommendation to weak accept.
Review #2
- Please describe the contribution of the paper
This paper proposes a framework that represents ECG signals using clinically meaningful waveform units called Cardio-Morphemes. By organizing these units into Cardio-Sentences, the method better captures the temporal and structural relationships in ECG signals. Based on this idea, the authors demonstrated improved downstream task performance and clinical interpretability. Overall, the paper is well-written and well-motivated.
- 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 experimental section is solid. The paper provides sufficient evidence of the predictive improvements achieved by the proposed framework through downstream evaluations on two datasets. 2) Overall, the figures are of high quality and well presented. 3) The concept of Cardio-Morpheme is interesting. It naturally equips the model with the ability to attend to clinically meaningful regions.
- 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 implementation details could be further elaborated. For example, the authors could provide more information about the training setup for the downstream tasks, as well as the number of transformer encoders used. 2) The paper does not mention whether the random seed was fixed or whether repeated experiments were conducted. If not, it would be better to run the experiments multiple times and report statistical significance tests against the baselines. For example, the AUROC improvement of ShapeEFM (89.90) over HeartLang (89.76) under the full-data setting may not be statistically significant. Including p-values would strengthen the robustness of the results. 3) The current discussion lacks depth. If possible, the authors could strengthen Section 3.3 by providing more clinically meaningful analyses, rather than only showing that similar patterns are clustered together. For example, they could analyze the statistical correlation between the learned features and clinical markers (e.g., the QRS complex), or show whether the model’s attention overlaps with clinically meaningful ECG segments.
- 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 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?
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 overall presentation of the paper is clear. The proposed framework improves downstream task performance while also providing clinical evidence, rather than being merely incremental. However, if the discussion can be further strengthened, I would be inclined to support acceptance.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Although the experiment part is more solid after rebuttal, the novelty of this paper is limited. Additionally, I still wish more discussion on the clinical perspective if the paper were to be accepted in MICCAI
Review #3
- Please describe the contribution of the paper
The paper proposes a tokenizer that generates the ECG sentences from the raw ECG signals based on individual waveform components, such as the P, Q, R, S, and T waves.
- 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.
In the EchoNext dataset, the paper attains the AUPRC of 5.04% at 100% labels.
- 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 method of the paper is similar to ‘Reading your heart: Learning ecg words and sentences via pre-training ecg language model’.
- 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
In Fig. 3(left), the yellow segments correspond to the ST seg, but it is incorrectly labeled as the TP seg.
- 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 experimental result is good. In the CSN and EchoNext dataset,it attains excellent AUROC and AUPRC at 10% and 100% labels.
- 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
Author Feedback
AC/R1/R2/R3: This response addresses key clarifications regarding method novelty, PIP segmentation robustness, the rationale for Lead-I-only segmentation, implementation details, statistical significance, and clinical interpretability. We revised the manuscript to make ShapeEFM’s novelty boundary clearer. HeartLang and ShapeEFM both use ECG language modeling, but with different levels of tokenization and self-supervised reconstruction targets. Unlike HeartLang, which compresses each beat into one word and largely relies on the accurate detection of the QRS complex, ShapeEFM treats sub-beat morphology changes such as QRS widening and ST deviation as reconstruction units, making these clinically relevant transitions explicit targets of SSL rather than implicit within heartbeat tokens. And these morphology-level shapelets enhance model’s representation capability for ECG morphology. Our PIP extraction algorithm relies on local waveform to reduce baseline drift’s effect, and we employ a low-pass filter to mitigate high-frequency noise. In addition, the codebook learned in the VQ step helps absorb subject-specific muscle artifacts after segmentation. Internal validation on the full MIMIC cohort using synthetic noise (±0.5mV baseline drift; 20-100Hz EMG) reveals that mean PIP location deviation was 1.2ms (less than 10% of the 16ms Cardio-Morpheme scale), supporting its stability under these perturbations. Our choice of Lead-I-only PIP segmentation leverages Lead I’s higher SNR relative to precordial leads, and provides a shared temporal scaffold for synchronized 12-lead ECGs. This strategy not only enhances robustness but also yields better performance: our internal ablation study comparing two segmentation strategies revealed that Lead-I-only segmentation achieved an AUROC (CSN 100%) of 89.80%, surpassing the 88.45% obtained by the multi-lead approach. For implementation, polyphase filtering is used to upsample EchoNext data from 250 Hz to 500 Hz, aligning it with pretraining frequency. ShapeEFM consists of 12 Transformer encoder blocks, 8 attention heads, and a dropout rate of 0.1, which is added to the manuscript. We commit to release the full codebase upon paper acceptance to facilitate reproducibility. We also conducted repeated experiments with 5 random seeds, and employed paired t-test to confirm ShapeEFM’s improvements are statistically significant. Notably, these gains are robust in multiple scenarios, including low-data regimes (CSN 10% AUROC: p=0.008) and full-supervision settings (EchoNext 100% AUROC: p=0.003). The statistical analysis have been incorporated into the revised Table 2.For clinical interpretability, we divided patients into ST-depression (>0.5mm) and normal groups, clustered the Cardio-Morpheme vocabulary into 12 categories, and analyzed distribution differences via the Chi-Square Test (p<0.05). Results indicate that the most significant cluster linked to ST-depression exhibits negative slopes or reduced amplitudes compared to normal levels (updated in revised Fig 3).
R1: We agree that our segmentation wording was too strong. ShapeEFM does not eliminate discrete segmentation, it mitigates fixed-window/heartbeat-level tokenization limits through adaptive morphology-aware segmentation. We have softened this claim in the manuscript. For Cardio-Morpheme configurations, both the PIP count and maximum length are fixed when the signal duration and sampling frequency are constant. We apply upsampling to accommodate inconsistent frequencies. We also fixed the symbol reuse issues.
R2: The association between specific clusters and ST-depression has been detailed in the preceding section. Regarding the suggestion to overlay model attention on clinically relevant segments, we agree this is also valuable and plan to pursue it as future work.
R3: The similarity to HeartLang is addressed above by clarifying the different morphology-level tokenization design and SSL reconstruction target. We fixed the Fig. 3 label error
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.
I recommend an Invite for Rebuttal. The reviewers praised the paper’s clear motivation to capture continuous ECG waveform morphology, the completeness of the technical pipeline, and the strong downstream empirical results on the EchoNext and CSN datasets. However, several critical technical and methodological concerns must be addressed. First, the authors must clarify the boundary of their key innovation, as reviewers noted the framework heavily resembles existing models like HeartLang, with the primary difference being the segmentation granularity rather than a fundamental departure from discrete segmentation. Second, the authors need to rigorously validate the robustness of the PIP-based segmentation algorithm against baseline drift, high-frequency noise, and muscle artifacts, while also justifying the strong assumption of applying Lead I’s segmentation boundaries across all 12 leads. Third, the rebuttal must supply missing implementation details, specifically regarding how inconsistent sampling frequencies across pretraining and downstream datasets were handled, the training setups, and the number of transformer encoders used. Furthermore, the authors must report statistical significance tests (e.g., p-values) for their performance improvements to ensure the reported gains over baselines are robust. Finally, the authors should deepen the clinical interpretability analysis by correlating learned features with established clinical markers, correct the mislabeled ST segment in Figure 3, and fix any inconsistent mathematical symbols in the methodology section.
- 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 introduces an ECG foundation model that captures sub-beat waveform morphology, and the authors’ comprehensive rebuttal effectively resolved initial concerns by clarifying the method’s novelty over existing baselines, proving the segmentation’s robustness to noise, and providing necessary statistical significance tests. Although one reviewer ultimately felt the methodological novelty remained limited, the thorough empirical validation, proven downstream performance, and the newly added clinical interpretability analyses linking learned features to ST-depression justify its 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.
Reject
- Please justify your recommendation.
The motivation is clear and important, while the empirical gains are incremental against strong baselines, and the evaluation is limited to linear probing on two downstream datasets, so the “foundation model” framing is somewhat stronger than the demonstrated transfer scope.
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.
Thank you for the detailed rebuttal. We appreciate your time and effort.
The rebuttal successfully addressed the main concerns raised during review. The authors clarified the novelty relative to prior ECG language models, repositioning ShapeEFM as a morphology-aware tokenization and reconstruction framework rather than a departure from discrete segmentation. Additional evidence on segmentation robustness, Lead-I design choices, implementation details, statistical significance, and clinical interpretability substantially strengthened the submission.
Although the novelty remains incremental and further clinical discussion would improve the work, the contribution is technically sound, reproducible, and supported by stronger experimental evidence.
I recommend to Accept this paper.