acceptodds
Withdrawn from ICLR 2027

ECGEchoDiag: A Benchmark for ECG-Conditioned Multi-View Echocardiography Generation and Cardiac Disease Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Electrocardiography (ECG) enables prolonged cardiac monitoring but cannot directly visualize cardiac anatomy or motion. ECG-conditioned echocardiography (Echo) generation offers a potential way to infer this information, yet whether generated videos preserve complementary diagnostic evidence across views remains unclear. We introduce ECGEchoDiag, a dataset and benchmark for ECG-conditioned multi-view Echo video generation and evaluation of diagnostic information preservation. It contains 15,715 synchronized ECG–Echo cardiac cycle pairs from 991 patients across 10 views, including 507 patients with multi-label annotations for eight cardiac disease categories. The dataset preserves within-patient associations across views and temporal synchronization within each recording. We further propose EchoGenDiag, which combines target-aware cross-view conditioning, cardiac-phase-dependent motion modeling, and reference-guided video reconstruction. For each target view, the model integrates its reference frame and ECG with information from other available views of the same examination. A frozen diagnostic teacher provides multi-label supervision and prediction consistency constraints to encourage disease-related information preservation. We evaluate reconstruction quality, temporal consistency, and patient-level multi-label disease recognition using patient-disjoint splits. Comparisons between real and generated videos assess diagnostic information preservation, while single-view and multi-view comparisons examine the contribution of cross-view conditioning.

Primary area:
datasets and benchmarks
Keywords:
ECG, Echocardiography, Multi-view, Video Generation, Cardiac Diagnosis, Multimodal Learning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_xyopOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces the ECGEchoDiag dataset and the EchoGenDiag method to generate cardiac-cycle videos from ECG and multi-view echocardiographic reference frames, and evaluates information preservation in the generated videos through disease recognition. Its main contributions are the synchronized ECG-Echo data with patient-level multi-label annotations and an evaluation that extends beyond visual reconstruction quality to disease recognition.

Critical strengths
  1. Synchronized ECG-Echo data with patient-level multi-label annotations.
  2. An evaluation that extends beyond visual reconstruction quality to disease recognition.
Critical weaknesses
  • The introduction emphasizes that when a patient’s cardiac condition changes between clinical visits, out-of-hospital monitoring often lacks imaging evidence of these structural and functional changes. The aim is therefore to infer cardiac structure and motion from ECG. However, the experiments still require a real echocardiographic first frame synchronously paired with the target ECG, creating a contradiction between the proposed application scenario and the task actually evaluated.

  • Table 3 compares ECG-only, first-frame repetition, and generated videos, but lacks a key baseline that directly feeds the same multi-view reference frames and ECG into a diagnostic model. Once ECG and echocardiographic first frames from multiple views are already available, why is it necessary to generate a video for diagnosis, and what problem does this process solve that direct prediction cannot?

  • The full method can access information from other views when generating the target view, whereas the multi-view baseline generates each view independently and subsequently aggregates diagnostic results. Thus, the comparison involves differences in both the information available during generation and the model architecture.

Questions for authors

Please see the weaknesses.

Reviewer_H8z7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ECGEchoDiag, containing 15,715 synchronized ECG–Echo cardiac cycles from 991 patients, including 507 patients with annotations for eight disease categories. It proposes EchoGenDiag, which combines ECG, target and support-view reference frames, and diagnostic-teacher supervision to generate multi-view Echo videos and evaluates reconstruction quality and patient-level disease recognition.

Critical strengths
  1. The dataset provides a useful combination of modalities and annotations. It covers ten Echo views, synchronized ECG, within-patient cross-view associations, and eight disease categories (Section 3 and Table 1), supporting research on diagnostic information preservation in generated videos.

  2. The experimental controls are reasonable. Table 3 includes real-video and first-frame repetition baselines, while Figure 5 separately removes disease supervision, ECG conditioning, and cross-view learning. The repetition baseline also illustrates why image-distribution metrics alone are insufficient to assess motion or diagnostic information.

Critical weaknesses
  1. The diagnostic evaluation protocol needs clarification. Equation 1 retains the real target reference frame, I^0t=I0t\hat{I}_0^t = I_0^t, and the full model also uses same-examination support-view reference frames. Table 3 reports AUPRC/AUROC of 0.518/0.783 for multi-view first-frame repetition and 0.590/0.821 for the full model. Section 5.1 describes the differences in generator conditioning across methods, but the exact inputs to the diagnostic evaluator—including whether the observed reference frame or ECG is included—are not sufficiently specified. This makes the diagnostic results and the contribution of generated future frames difficult to interpret.

  2. Some baseline results and key method definitions are missing. EchoDiffusion-FT, EchoFlow-FT, and ECHOPulse-ZS are described in Section 5.1 but absent from Table 3, without explanation. Sections 4.2–4.3 also leave the construction or dimensions of dd, Δ\Delta, gmg_m, gctxg_{\mathrm{ctx}}, and s~t\tilde{s}^{t} insufficiently specified. These quantities determine the central conditioning pathway, so the missing details affect reproducibility and verification of the comparisons.

Questions for authors
  1. What exactly does the diagnostic evaluator receive? Are results available for generated future frames with the observed reference frame excluded and support-view availability matched?

  2. Were the three omitted baselines evaluated, and why are their results absent? How are the cross-view representations, residuals, gates, and state variables constructed?

Reviewer_zUK9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper releases ECGEchoDiag, a single-center dataset of 15,715 synchronized ECG-echo cardiac cycles from 991 patients over 10 views, with patient-level labels for 8 disease groups on 507 patients. It also proposes EchoGenDiag, which generates a 21-frame, 128x128 echo cycle from a real first frame, ECG and first frames of other views, trained with a frozen diagnostic teacher (BCE on generated videos plus teacher consistency). Evaluation uses FID/FVD/Sector-SSIM and macro AUPRC/AUROC of a frozen classifier, compared with first-frame repetition and fine-tuned ECHOPulse.

Critical strengths
  1. A dataset that pairs synchronized ECG with multi-view echo and disease labels, with patient-disjoint splits, does not really exist publicly right now. If it is actually released, it would be useful for the community.
  2. The paper asks the right question: does a generated echo keep diagnostic information, not just look similar? The finding that first-frame repetition gets a good FID with zero motion is a useful warning about how this area is usually evaluated.
Critical weaknesses
  1. The main claim is not supported by the results. In Table 3, simply copying the real first frame of each view (no motion at all) already gets AUROC 0.783 in the multi-view setting, versus 0.821 for the full model and 0.830 for real videos. So most of the diagnostic signal comes from the real input frames, not from the generated motion. On top of that, the generator is trained with BCE against the disease labels through a frozen teacher, and then evaluated by a frozen classifier, and the paper never says whether these two share architecture, data or weights. So the AUROC may partly measure "the model learned to please the classifier" rather than real diagnostic content. With 481 of 507 labeled patients positive, about 148 test patients, some classes with very few cases, and no confidence intervals or per-class results, a 0.04 AUROC gap over the trivial baseline is not convincing.
  2. The submission looks unfinished. The text block is clearly wider than the ICLR template (x = 72-540 pt vs 108-504 pt in the standard style), which fits noticeably more content into the page limit. Table 3 uses internal run names ("V21 Hybrid-Delta", "Ours-SV V21 strict A4C"); EchoDiffusion-FT, EchoFlow-FT and ECHOPulse-ZS are listed as baselines but do not appear in any table; the target-aware cross-view fusion, which is the main new module, is never actually defined in Section 4, and neither is L_base; the appendix is just the template sentence "You may include other additional sections here."; and "temporal consistency" is claimed as an evaluation axis but no temporal metric besides FVD is reported.
Questions for authors
  1. Is the frozen evaluation classifier the same network as the diagnostic teacher used in training? If not, how do they differ in architecture, training data and initialization?
  2. Can you evaluate with the first frame removed from the evaluator input, or compare against a "real first frame + mean motion" baseline, to show how much the generated motion adds?
  3. Please report per-class AUROC/AUPRC with confidence intervals (e.g. patient-level bootstrap) and results over several patient-level splits.
  4. Is the dataset really going to be public? Where, under what license, and through what access process? The data come from DICOM files with patient information, so how exactly was de-identification done and checked?
  5. What is the agreement between the two echocardiographers on end-diastole and cycle extraction? And is there any external test set (other hospital or vendor), since all data are from one hospital and Philips machines?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.