List of Papers Browse by Subject Areas Author List
Abstract
This study introduces a novel demographic-conditioned architecture that leverages Mamba2 State Space Models (SSMs) and a dual Attention Aggregation mechanism to integrate 12-lead ECGs with patient sex metadata for biological age estimation (ECG-age). The architecture efficiently captures long-range physiological dependencies, while an attention-based aggregation mechanism renders the model invariant to lead availability and signal duration. Trained on the full CODE cohort (N=1.3 million) and validated on the CODE-15% test set (N=233k), the model achieves a mean absolute error of 6.72±5.81 years (R^2=0.80) in chronological age prediction. This performance surpasses a common used 1D-ResNet baseline (8.44±7.12 years, R^2=0.69). This improved precision strengthens ECG-age as a prognostic biomarker: patient hearts predicted to be over 8 years older than their chronological age exhibited a higher Hazard Ratio (HR) for all-cause mortality (HR 2.06, 95% CI 1.92-2.21) compared to the 1D-ResNet baseline (HR 1.80, 95% CI 1.68-1.92) on the CODE-15% test set. These results demonstrate that the SSM-based architecture yields an ECG-age marker that is prognostically superior to current baseline models. Code is available at: https://github.com/B-Bracke/demographic-ssm-ecg-age
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3670_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/B-Bracke/demographic-ssm-ecg-age
https://huggingface.co/BBracke/demographic-ssm-ecg-age
Link to the Dataset(s)
CODE dataset: https://figshare.scilifelab.se/articles/dataset/CODE_dataset/15169716
CODE-15% dataset: https://zenodo.org/records/4916206
Sami-Trop dataset: https://zenodo.org/records/4905618
BibTex
@InProceedings{BraBen_DemographicConditioned_MICCAI2026,
author = { Bracke, Benjamin AND Stang, Andreas AND Schmidt, Börge AND Friedrich, Christoph M.},
title = { { Demographic-Conditioned State Space Models for ECG-Based Age Estimation } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16896},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper proposes a multimodal state space model for predicting biological (ECG-derived) age from electrocardiograms. The model employs a convolutional stem followed by a bidirectional Mamba2-based backbone with adaptive layer normalization and dual-axis attention aggregation. The framework incorporates biological sex metadata for conditioning and is evaluated on a large-scale ECG dataset (CODE, N≈1.3M), with external validation on SaMi-Trop. The study compares sex-conditioned and sex-agnostic variants and evaluates downstream associations with mortality risk using hazard ratios.
- 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 leverages a very large-scale dataset (CODE), which is a significant strength and provides strong statistical power.
- The proposed architecture adopts modern sequence modeling techniques (state space models and attention-based aggregation), which are relevant for long ECG signals.
- The inclusion of external validation (SaMi-Trop) and survival analysis (hazard ratios based on ECG-age deviation) goes beyond simple regression metrics and connects the task to clinical outcomes.
- The comparison between sex-conditioned and sex-agnostic modeling provides an interesting and practically relevant perspective.
- 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.Limited comparison with more recent approaches While there is no clearly established state-of-the-art benchmark for ECG-age prediction, comparisons are primarily limited to an earlier baseline (i. e. , the 2021 1D-ResNet-based model). As a result, it is difficult to assess how the proposed method relates to more recent advances in ECG representation learning.
2.Generalization limitations on external data Although the proposed model shows improved performance and stronger risk stratification on the CODE dataset, the generalization to external data (SaMi-Trop) is limited, with relatively low R² values (0.35–0.38) and lack of statistically significant hazard ratio results. This weakens the overall robustness of the findings.
3.Limited strength of clinical impact The paper includes discussion on prognostic value and demonstrates improved risk stratification compared to the baseline. However, the overall impact remains somewhat limited, as the improvements are not consistently supported across datasets, particularly in external validation.
4.Incremental problem formulation The work follows the established paradigm of ECG-based biological age estimation. While the proposed architecture incorporates modern modeling techniques, the overall problem formulation remains similar to prior work, limiting the novelty of the contribution.
- 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?
The paper presents a well-executed study with a modern modeling approach and leverages a very large dataset, which is a significant strength. The problem is relevant, and the inclusion of survival analysis adds clinical context. However, the overall contribution is somewhat incremental, and it remains unclear whether the observed improvements in ECG-age prediction translate into meaningful clinical benefits. The limited comparison with recent methods and reduced generalization on external data further weaken the claims. Despite these limitations, the scale of the study and the engineering contribution make it a valuable addition, and it falls slightly above the acceptance threshold.
- 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.
After reading the rebuttal, I maintain my original Weak Accept-level recommendation. The authors reasonably clarified the rationale for the baseline and dataset choices, but the lack of comparisons with recent Transformer/SSM-based ECG models and the weaker external validation results remain limitations of the submitted paper. Overall, the paper is still slightly above the acceptance threshold due to its large-scale evaluation, modern modeling approach, external validation, and survival analysis.
Review #2
- Please describe the contribution of the paper
The main contribution of the paper is the development of an ECG-based age prediction model built on a State Space Model (Mamba) backbone with a dual attention aggregation mechanism. The goal is to estimate ECG-based (biological) age from 12-lead ECGs, optionally together with patient sex metadata. The model is trained on a big cohort of ECGs and shows improvement over baselines.
- 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 paper is written in clear language and is easy to follow.
2.The motivation and idea are clear and well-explained. Predicting biological age from ECG is, in my view, a strong and interesting idea with potential clinical relevance.
3.The use of a large dataset makes the results more reliable and increases confidence in the overall findings.
4.The observation that sex-agnostic prediction performs similarly to, or comparably to, sex-conditioned prediction is valuable.
5.All implementation details are provided, making this work reproducible.
- 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 claim of methodological novelty is not fully convincing, as related ECG works using Mamba/SSMs (e.g., ECGMamba, S2M2ECG) exist and are not discussed. Even if recent, they should at least be referenced.
2.The choice of baselines is limited. In particular, a Transformer/ViT-style baseline (e.g., convolutional stem + transformer backbone), which is now common in ECG modelling, is missing.
3.The choice of 400 Hz sampling rate is not justified, especially since 100 Hz and 500 Hz are more standard in the literature.
4.While the paper mentions interpretability via temporal and lead attention, no qualitative analysis is shown. In my view, once it is included in the figure and mentioned in the text, it is important to demonstrate which patterns the model attends to and whether they are clinically meaningful.
5.Regarding the reported improvements: while the model reduces MAE (e.g., from ~8.4 to ~6.7 years), it is unclear whether this difference is clinically meaningful. For me, it is questionable whether distinguishing between, for example, a 9-year vs. 7-year age gap would materially affect clinical decision-making.
6.The use of the term “multimodal” may be somewhat overstated, as the model incorporates only a single scalar metadata variable (sex) alongside the ECG signal, rather than multiple high-dimensional or semantically distinct modalities.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
Mentioned references: 1.ECGMamba: Towards Efficient ECG Classification with BiSSM - https://arxiv.org/abs/2406.10098 2.S2M2ECG: Spatio-temporal bi-directional State Space Model Enabled Multi-branch Mamba for ECG - https://arxiv.org/abs/2509.03066 3.MSECG: Incorporating Mamba for Robust and Efficient ECG Super-Resolution - https://arxiv.org/abs/2412.04861
- 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?
Thanks to the authors for the work.
I believe the work is well motivated and built around a strong idea. The paper is easy to follow and clearly written. The results appear reliable, supported by training on a large-scale cohort. I found the observation that the raw ECG signal may already capture sex-related physiological differences both interesting and potentially clinically relevant.
The main weaknesses are the limited set of baselines and the related work discussion. In particular, the lack of comparison with a ViT/Transformer-based approach is a notable omission, given its current relevance in ECG modelling. Additionally, similar recent works using Mamba/SSM architectures are not referenced or discussed, which weakens the novelty claim.
- 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 author’s feedback partially addresses my concerns, and I would keep a positive recommendation.
The authors clarify the choice of baseline, the 400 Hz sampling rate, and the use of the term “multimodal,” which I find reasonable.
However, I still think the clinical utility of the MAE improvement is not fully resolved. The rebuttal states that MAE decreases from 8.44 to 6.72 years, with reduced variance and improved hazard ratios. This supports better statistical performance, but it still does not fully explain how clinical decision-making would change in practice. For example, it remains unclear whether distinguishing between a predicted ECG-age gap of approximately 7 years versus 8–9 years would lead to different clinical actions. I therefore think the paper should frame ECG-age mainly as an adjunctive prognostic or screening signal, rather than implying direct clinical actionability.
The authors state that the main difference from prior ECG SSM models is the dual attention aggregation mechanism. However, this is not analyzed in the paper. Without isolating the contribution of the attention module, it is difficult to know whether the performance gain comes from this component specifically, from the backbone, from dataset scale, or from other implementation choices.
Overall, despite these limitations, the work is well executed, uses a large-scale dataset, and presents useful results. I would maintain my Weak Accept recommendation.
Review #3
- Please describe the contribution of the paper
o Proposes a novel multimodal architecture for ECG-based biological age estimation that combines convolutional feature extraction with bidirectional Mamba 2 state space models to capture long-range temporal dependencies in 12 lead ECG signals (fusing ECG features with biological sex) o Integrates patient metadata (sex) via adaptive layer normalisation, and evaluates both sex-conditioned and sex-agnostic inference strategies o Trains and evaluates the model at scale on the large (1.3M recordings) CODE cohort, with internal validation on CODE 15 and external validation on the SaMi Trop cohort o Demonstrates improved age estimation over a 1D ResNet baseline
- 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.
o Well written and well structured o Proposes a novel and well-articulated architecture combining bidirectional state space models with attention-based aggregation for long ECG sequences o Strong technical motivation for architectural choices, which are all clearly explained step by step o Trained and evaluated on a very large-scale dataset with additional external validation, which strengthens the empirical credibility o Demonstrates consistent improvements over a commonly used 1D‑ResNet baseline for age prediction and prognostic stratification
- 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.
o Clinical utility is insufficiently justified. The model’s MAE (~7 years) is close to the 8‑year ECG‑age threshold used for prognostic stratification. This raises concerns about misclassification across the ±8 year boundary, and potential overinterpretation in clinical settings, purely due to prediction noise. In practice, this high error may over- or under-alarm clinicians/patients o The term “multimodal” appears overstated, as the only non‑ECG input is biological sex, which provides limited additional information and appears redundant in some experiments o While conceptually interesting, the high prediction error limits near-term clinical applicability, reducing the relevance of the contributions for MICCAI’s translational focus o The discussion does not sufficiently address how ECG‑age predictions should be used responsibly in practice, given the observed uncertainty
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
Molecular profiling (e.g., proteomics) is limited currently, but it may well become more widely available as evidence builds supporting its use in organ age profiling, e.g., work by Tony Wyss-Coray’s team!
The experimental design is sound and includes large-scale training and external validation.
- 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 is a technically strong and thoughtfully designed study with a clear methodological contribution. However, the clinical value proposition is not yet convincing: the prediction error remains high relative to the prognostic threshold, the multimodal aspect is limited, and the practical implications of ECG‑age estimates are insufficiently discussed. As a result, the work appears promising but premature for acceptance at MICCAI in its current form.
- 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
General Response to All Reviewers: We sincerely thank the reviewers for their constructive feedback. We agree that further evaluations are vital. However, due to MICCAI’s strict 8-page limit, we carefully scoped this submission as a foundational proof-of-concept. What separates our approach from prior ECG SSM models is our unique dual attention aggregation mechanism, which ensures robustness against varying signal durations and missing leads, coupled with a flexible integration mechanism for tabular metadata. Specific concerns are addressed below. ——-
1.Baselines, Dataset Selection, and Future Cohorts (Reviewers 1 & 2)
Critique: Missing comparisons to Transformers and SSM-based models, and limited external generalization.
Response: We acknowledge the omission of recent parallel SSM works and will add references to ECGMamba, S2M2ECG, and MSECG in the revision. We selected the 1D-ResNet as our primary baseline because it is the seminal architecture published alongside the CODE and SaMi-Trop datasets, enabling a direct and rigorous comparison on the exact same data splits. Furthermore, validating the prognostic value of ECG-age requires large-scale datasets linked with long-term mortality data, which exclude public sets like PTB-XL. Because SaMi-Trop is a severe-pathology cohort (Chagas cardiomyopathy) with double the mortality rate of the training data, a domain gap and performance drop are expected. To address this and conduct broader validations, we are applying for data access to the Harvard-Emory ECG Database (10M ECGs), the SHIP and HNR cohorts (Germany), and MIMIC-IV. Extensive benchmarking against recent Transformers and SSMs on these diverse cohorts is underway for a journal extension.
2.Clinical Utility vs. MAE (Reviewers 2 & 3)
Critique: Is a drop in MAE clinically meaningful given the 8-year threshold? Could this cause misclassification?
Response: We understand the concern regarding prediction noise near the 8-year threshold. However, the MAE reduction (8.44 to 6.72) is coupled with a substantial variance reduction (SD dropped from 7.12 to 5.81). This tighter error distribution means fewer patients are misclassified across the ±8-year boundary due to random noise. This precision drives our statistical improvement in risk stratification (HRs for all-cause mortality improved from 1.80 to 2.06). We will clarify that ECG-age is intended as an adjunctive screening flag in routine exams, not a standalone diagnostic, mitigating false alarm risks
3.The “Multimodal” Label (Reviewers 2 & 3)
Critique: “Multimodal” is overstated since the only non-ECG input is biological sex.
Response: We agree that utilizing only biological sex limits the traditional scope of “multimodal.” Our intention was to propose a foundational architecture capable of deep metadata fusion that can scale with high-dimensional metadata (e.g., extensive comorbidities) in future work. We will tone down the “multimodal” claims in the text and are open to adjusting the title (e.g., “Demographic-Conditioned”).
4.Technical Specifics: 400 Hz and Explainability (Reviewer 2)
Critique: Why a 400 Hz sampling rate, and where are the attention visualizations?
Response:
- 400 Hz Sampling Rate: While 500 Hz is standard clinically, 400 Hz is the native preprocessing standard established by the CODE dataset authors. To maintain strict comparability with the 1D-ResNet baseline, we adhered to the same dataset-specific sampling rate.
- Explainability: We agree visual explainability is crucial. However, merely overlaying attention weights onto ECG signals without rigorous clinical context can be misleading. Proper analysis requires carefully linking scores to established physiological ECG features. Due to the 8-page limit, we lacked the necessary space for this evaluation. A comprehensive interpretability analysis is a core component of our planned extended journal submission
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.
The reviews are mixed, with two weak-accept recommendations and one weak-reject recommendation. The paper presents a SSM/Mamba-based architecture for ECG-based age estimation, trained on a large CODE cohort and evaluated with both internal and external validation. The reviewers recognize several strengths, including the scale of the study, clear architecture, improved regression performance over a 1D-ResNet baseline, sex-conditioned versus sex-agnostic analysis, and the inclusion of survival analysis for prognostic relevance.
However, the reviewers raise several important concerns. The main issues are: (1) the methodological novelty is incremental, and related recent ECG Mamba/SSM and Transformer-style baselines are not sufficiently discussed or compared; (2) the external validation results are weaker than the internal results, with limited R2 and less consistent prognostic significance on SaMi-Trop; (3) the clinical utility of the ECG-age gap remains insufficiently justified, especially given that the MAE is close to the 8-year prognostic threshold; and (4) the term “multimodal” may be overstated because the only non-ECG input is biological sex, and the interpretability claims based on attention are not yet supported by qualitative analysis.
- 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 considering the reviews and rebuttal, I recommend acceptance.
The rebuttal reasonably clarifies the baseline choice, the use of the CODE-standard 400 Hz sampling rate, the expected domain gap on SaMi-Trop, and the intended role of ECG-age as an adjunctive prognostic or screening signal rather than a standalone clinical decision tool. The authors also agree to tone down the “multimodal” framing. Important limitations remain, including incremental novelty, lack of recent Transformer/SSM ECG baselines, weaker external validation, and limited evidence for clinical actionability. Nevertheless, given the scale, reproducibility, external evaluation, and generally positive reviewer opinions, the paper is above the acceptance threshold.
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.
Although evaluation with newer methods is mentioned, the reviewers concur that this is a paper with a comprehensive evaluation. The survival analysis is relevant for a clinical application of age estimation. My assessment is that the positive aspects outweigh the weaknesses, so I suggest 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.
This is an interesting and clinically relevant study, particularly in using ECG-age for mortality risk stratification and testing generalization in a severe-pathology cohort. Although the “multimodal” claim should be toned down and broader SSM/Transformer baselines would strengthen the work, the rebuttal reasonably addresses the main concerns, and the reported improvements in MAE, variance reduction, and hazard ratio suggest meaningful potential clinical value.