List of Papers Browse by Subject Areas Author List
Abstract
Electrocardiogram (ECG) foundation models (FMs) are increasingly deployed as cloud services. For downstream tasks such as disease diagnosis, they return compact embeddings from input ECG signals. In this paper, we find that such embeddings pose significant privacy risks: adversaries can use the exposed embeddings and invert the original ECG, which leads to the inference of sensitive clinical attributes. To assess this vulnerability, we introduce SHINE, a new ECG inversion method that inverts ECG signals from the exposed embeddings. SHINE consists of ECG mapper and ECG inverter. Specifically, SHINE trains ECG mapper to approximate the embedding space of victim ECG FM, which is then used to synthesize the necessary data to train ECG inverter from scratch. Our evaluation against four state-of-the-art ECG FMs shows that SHINE achieves high-fidelity inversion. Importantly, the inverted ECGs retain sensitive clinical attributes, including cardiovascular disease indicators up to 99.6%, almost identical to the original ECG. These results demonstrate that the exposed ECG embeddings preserve substantial physiological information so that when deploying ECG FM, stronger security countermeasures need to be devised. To our knowledge, this paper is the first to report that the exposed embeddings can be inverted to the original ECG signals.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2919_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
PTB-XL: https://physionet.org/content/ptb-xl/1.0.3/
CPSC2018: http://2018.icbeb.org/file/2018X_Feifei_An%20Open%20Access%20Database%20for%20Evaluating%20ECG%20abnormal%20classificaition%20algorithm.pdf
BibTex
@InProceedings{LeeJun_Inversion_MICCAI2026,
author = { Lee, Junseok AND Jung, Seungwoo AND Yang, Gyeongsik AND Yoo, Chuck},
title = { { Inversion from ECG Foundation Model } },
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
1.The paper is the first to demonstrate that embeddings from ECG foundation models can be inverted back into clinically meaningful ECG signals.
2.It presents the SHINE method, a query-efficient two-stage attack framework that operates within realistic cloud constraints, comprising an ECG Mapper and an ECG Inverter. It overcomes two key barriers: the absence of pretrained ECG decoders and the limitations imposed by query budget constraints.
- 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 the first to frame ECG embedding inversion as a privacy attack vector against cloud-deployed foundation models, which is novel. 2.The paper presents a clever use of pseudo QR pairs, in which rather than requiring massive real-world queries to the victim model (which is infeasible under cloud constraints), SHINE trains a lightweight ECG mapper on a small number of real query-response pairs and then uses it to generate a large synthetic training set. 3.The evaluation is notably thorough across multiple dimensions, testing against both CNN-based and transformer-based architectures, but also incorporating clinically meaningful interval metrics such as the absolute differences between RR intervals and QT durations extracted from the original and inverted ECGs.
- 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 chosen baseline (adapted DiffuSETS) is acknowledged by the authors themselves to be a poor fit. Stronger baselines should be considered, such as gradient-based optimization in embedding space or nearest-neighbor retrieval, to better demonstrate SHINE’s true advantage. 2.SHINE is trained and evaluated on PTB-XL and CPSC2018 only. Evaluation on more diverse datasets (e. g. , MIMIC-IV-ECG) would strengthen the generalizability claims. 3.The paper calls for stronger countermeasures but evaluates none. Even a basic analysis of how SHINE performs under differential privacy noise or embedding perturbation would significantly strengthen the practical impact of the work.
- 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.
(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 addresses a genuinely novel and clinically important problem, demonstrating for the first time that ECG foundation model embeddings can be inverted to recover sensitive patient information. The pseudo QR pair synthesis strategy is technically clever and the evaluation is reasonably thorough. These are meaningful contributions that warrant publication. However, several weaknesses prevent a stronger score:
1.The baseline comparison is too weak to convincingly demonstrate SHINE’s true advantage, and stronger baselines (gradient-based optimization, nearest-neighbor retrieval) are notably absent. 2.The absence of any defense evaluation limits the practical impact of the work. 3.The lack of a discussion section and the indirect clinical evaluation leave important questions unanswered
The paper is marginally above the acceptance threshold primarily because the problem formulation is novel and important for the MICCAI community, and the results are compelling enough to provoke meaningful discussion about privacy in clinical AI deployments. However, the authors should be expected to address the weak baseline comparison and the absence of defense evaluation 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
Review #2
- Please describe the contribution of the paper
The contribution of the paper consists of two main models. The first one is an ECG mapper, which aims to approximate and learn the transformation between two spaces, mapping ECG signals into the embedding space of the target model. The second one is an ECG inverter, which is trained using pseudo-synthetic data generated by the mapper, where real ECG signals are paired with embeddings produced by this proxy model. Using these pairs, the inverter learns to reconstruct the original ECG signal from the embedding representation.
- 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.
- Query-efficient training strategy: The use of a surrogate mapper to approximate the embedding space reduces the need for extensive query access to the victim model, which can ultimately be related to the optimal subset of data that represents the total variance in a dataset (a data distillation idea).
- Clinically meaningful evaluation: The study evaluates not only signal reconstruction but also the preservation of clinical attributes, strengthening the practical relevance of the results, arguably the most important aspect. The authors cite [23], whose core idea suggests that this approach can serve as insight for black-box interpretability during clinical decision-making.
- Relevant and timely problem: The paper addresses privacy risks in ECG foundation models, particularly in cloud-based settings. It is a relevant idea; in my opinion, however, it does not fully capture the potential of this work. This aspect of privacy becomes a necessity when most of these foundation models are commercial and potentially used in real clinical decision-making.
- 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 proposed approach mainly consists of combining standard components (CNN/Transformer architectures and MSE loss) within a training pipeline. The novelty lies primarily in the use of a surrogate mapper and pseudo pairs, which may be seen as an incremental contribution rather than a new method base don MIC outlines.
- The performance of the inverter is heavily dependent on the quality of the learned mapper. Any approximation error in the mapper may propagate to the pseudo query-response pairs, potentially affecting the fidelity of the inversion.
- Although the paper is framed as a privacy attack, the main contribution is methodological (training strategy), raising questions about whether the work truly advances understanding of privacy risks or mainly proposes a reconstruction pipeline. This continue into the lack of discussion on alternative interpretations, which basically, the obtained results could also be interpreted from a representation learning or interpretability perspective (i.e., what information is retained in embeddings), but this angle is not explore
- Maybe a crucial idea is the dataset bias and distribution alignment, where the proposed ethod relies on public datasets (e.g., PTB-XL) to approximate the embedding space. But never used other types of dataset like MUSIC or the demo of MIMIC-IV, or what happens the real clinical data were several methods are constructed to digitalize the ECG, where exist a shift-domain that shows when a ECG is digitalized how the leve lof signal (Pixel-mV) affects the reconstruction or diases identificators. This idea affects the robustness of the method under domain shift (e.g., different hospitals, devices, populations) is not evaluated.
- 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
I appreciate the work, and the underlying idea is crucial. However, in my opinion, this moment calls for stronger explainability and interpretability of large black-box models, such as foundation models (FMs).
- 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?
Based on the MIC evaluation, this paper should be rejected. However, the real potential of this work lies in interpretability, specifically, in how a function can reconstruct an ECG signal capable of identifying disease-related patterns (although this must be validated on real-world datasets, where domain shift may significantly affect performance). Furthermore, extending this approach to define thresholds for each condition represents a promising direction with substantial potential.
- 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
Review #3
- Please describe the contribution of the paper
This study investigates the privacy vulnerabilities of ECG foundation model embeddings, showing that exposed embeddings can leak sensitive information through inversion attacks. To overcome the challenge of limited query–response (QR) pairs, the authors introduce an ECG mapper that approximates the victim model, allowing the creation of pseudo QR pairs. These pseudo pairs, combined with real QR pairs, are used to train an ECG inverter. Experimental results demonstrate that ECG signals reconstructed from embeddings retain clinical prediction performance similar to the original ECGs, highlighting that the embeddings preserve significant clinically relevant information and present a substantial privacy risk.
- 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 is one of the first to examine privacy vulnerabilities in ECG foundation model embeddings, presenting a timely and well-justified problem.
- Introducing an ECG mapper to generate pseudo QR pairs offers a practical and novel solution to the challenge of limited query access in black-box settings.
- The method shows promising performance: inverted ECGs achieve clinical prediction results comparable to those of original ECGs, reinforcing the claim of significant information leakage.
- Evaluating the method on an external dataset strengthens the experimental design and further supports its generalizability.
- 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 authors do not clarify why the ECG mapper must use the same CNN-based or transformer-based architecture as the victim foundation model. This rationale should be evaluated, ideally through an ablation study.
- Metrics like MAE (e.g., 0.47) lack sufficient context, making it difficult to assess their relevance to meaningful differences in ECG waveform morphology. While RR and QT intervals are more interpretable, the clinical implications of the results are not addressed.
- The rationale for comparing against the proposed baseline is unclear, particularly if no strong or established baseline exists. Instead, the authors should focus on thoroughly justifying and evaluating each component of their method, as well as systematically varying model architectures. Demonstrating the performance of each choice under different real QR pair budgets would provide more meaningful insights.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper addresses a timely and important problem with a clear and novel formulation, and proposes a practical solution through the use of pseudo query–response pairs. The empirical results are convincing, particularly the demonstration that inverted ECGs retain clinically relevant information, which highlights a meaningful privacy risk. The use of an external dataset further strengthens the evaluation.
While there are notable weaknesses—particularly in the justification of architectural choices, interpretability of reconstruction metrics, and clarity of baseline comparisons—these issues primarily affect the depth of analysis rather than the core contribution. Overall, the paper makes a valuable contribution and merits 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.
The authors have clarified my concerns. I do not have any further comments.
Author Feedback
1.Experiment datasets[MR,R1,2,3] The main goal of our study is to demonstrate that exposed ECG embeddings can leak sensitive clinical attributes, which makes ECG FM vulnerable to privacy attack. Four FMs we test were trained with many datasets such as Harvard-Emory, MIMIC-IV-ECG, and CODE-15%, but not with PTB-XL dataset. So PTB-XL is a natural choice for us to ensure attacks from unknown data. Then CPSC2018 is used to evaluate our ECG mapper and inverter. Please note that CPSC2018 dataset has different distribution from PTB-XL so that we can confirm that the results are not driven by PTB-XL-specific bias. Furthermore, we tested with other dataset (CSN) and had similar results, but we could not include due to the space limit. 2.Limited defense evaluation[MR,R1] Yes, our current evaluation does not include defense methods. In privacy-attack perspective, we believe revealing the attack surface itself has merit. Our work follows other leakage analysis studies that revealing and characterizing practical vulnerabilities is treated as a primary contribution [Carlini et al., USENIX Security ’21]. We leave the evaluation of defense mechanisms to future work. 3.Baseline selection rationale—gradient-based optimization, nearest-neighbor retrieval[R1,3] Because this is the first to reveal ECG inversion, there is no well-established baseline. So we select a baseline from other domains (Vision) that are compatible with our threat model and query budget. We find that gradient-based optimization is impractical because it requires nearly 120k queries for a single inversion, far exceeding our ≤1000 budget. We did not choose nearest-neighbor retrieval as it retrieves the closest existing sample from a predefined candidate pool, rather than reconstructing the target input. Thus, we choose a pretrained decoder-based inversion as our baseline: it uses a decoder to reconstruct the input from exposed embeddings, following [DiffusETS, Patterns ’25]. 4.Clinically meaningful evaluation[MR,R1,3] The CPSC2018 dataset includes 9 diseases: 1st-degree AV block, atrial fibrillation, left bundle branch block, premature atrial contraction, right bundle branch block, sinus rhythm, ST depression, ST elevation, and ventricular ectopics. We ran experiments for all 9 diseases. Averaged over 4 ECG FMs and 2 direct QR pair budgets, SHINE achieves normalized AUROCs of 96.0% for atrial fibrillation and 95.7% for left bundle branch block, with similar results for other diseases. Due to the space limit, however, we present only the average across the nine diseases in Table 2.5.Robustness of direct/pseudo QR pairs[MR,R2,3] We ran experiments with a number of direct QR pairs between 250 and 8000.We find that SHINE can preserve sensitive attributes with 500 direct QR pairs. Specifically, SHINE achieves normalized AUROCs of 90.3% for age and 95.2% for disease over 4 FMs (§3.2). This indicates that SHINE approximates victim ECG FM’s embedding well enough by using only 500 direct QR pairs. With 500 direct QR pairs, we find that the right number of pseudo QR pairs for fidelity is more than 20000 (21337 in our case). 6.Architecture selection rationale[MR,R3] We ran various experiments to find out a proper model architecture for ECG mapper and inverter. The results show that the proper architecture depends on FM. In other words, if FM adopts CNN, the mapper and inverter of CNN gives the best approximation. So for the experiments with 4 FMs, we design ECG mapper and inverter, depending on FM. Regarding whether FM architecture can be known, there are some literature that identify the model architecture from their embeddings. 7.Generalization over diverse FM[MR] Yes, it is a next step. But, please note that ECGFounder has been trained with 10.7M ECGs, which makes it most extensive FM that exists today. So since our attack is feasible with only 500 queries for ECGFounder, we expect that our attack can be effective to other (small) FMs even with less than 500 queries.
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 SHINE to invert ECG foundation model (FM) embeddings back to original ECG signals. The approach is novel and valuable from a privacy-attack perspective. However, given the small datasets used, there are several key points the authors should address in a rebuttal:
1.Experiments are limited to PTB-XL and CPSC2018.The authors should discuss SHINE’s robustness across different hospitals, devices, populations, and sampling rates, and clarify the limitations of small datasets for foundation model applications. Robustness of pseudo QR pairs: SHINE relies on ≤1000 direct QR pairs to generate pseudo QR pairs. The authors should analyze how the quantity and quality of these pseudo pairs affect inversion fidelity, and indicate the minimum number of direct QR pairs required for stable performance. 2.The mapper and inverter are matched to the victim FM’s architecture (CNN or Transformer). The authors should clarify how SHINE performs if the architecture is unknown or misestimated, and how mapper errors propagate to the inverter. 3.More clinically meaningful metrics should be provided (e.g., arrhythmia detection, disease-specific sensitivity/specificity), and evaluation on additional heterogeneous datasets (e.g., MIMIC-IV-ECG) would strengthen claims. 4.SHINE’s performance under embedding perturbations, differential privacy, or other defenses should be discussed or simulated. 5.Authors should clarify the applicability and generalizability of SHINE to larger, more diverse ECG FMs and explain how small dataset experiments translate to foundation model settings.
- 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 addresses a novel and clinically important privacy and interpretability question: whether ECG foundation model embeddings can be inverted to recover sensitive signals. Reviewers note meaningful technical ideas and a reasonably thorough evaluation, with remaining concerns about stronger baselines and validation on broader real-world data. The overall assessment is positive.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The authors addressed Reviewer’s concerns. I would recommend to include it in camera-ready paper.
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 probably the most novel and consequential paper in my stack. It identifies a concrete privacy vulnerability of ECG foundation-model embeddings. The experiments are broad for the problem scope, covering four ECG foundation models and showing clinically meaningful leakage. All reviewers like the paper, and I personally like this paper very much.
But the authors shall not view this work as the final product of their project. It would be nice to see the systematic mitigation strategies in their potential journal extension. Best of luck and looking forward to seeing this work carry out the real contribution to the AI medical community!
