acceptodds
Under review as a conference paper at ICLR 2027

CARP: A Framework for Clinically-Aligned ECG Generation with World Models

20 views

Abstract

Electrocardiograms (ECGs) are recordings of the heart’s electrical activity over time. Text-to-ECG generation enables critical applications in rare-label data synthesis, robust model evaluation, and complex disease simulation. This generation problem is challenging since ECGs associated with clinically-distinct text largely share the same waveform structure and differ only in subtle morphological features and measurements. Recent diffusion-based generators produce plausible ECGs that are only partially clinically-aligned with the report, missing subtle morphologies and implied findings, particularly for rare conditions. To address these limitations, we introduce CARP, which adapts a pretrained world model to text-to-ECG generation and augments rectified-flow training with direct, sample-specific supervision over morphological measurements. Because these measurements are typically computed via non-differentiable waveform segmentation, we develop a novel differentiable soft-segmentation objective that aligns generated and target ECGs and enables training. Across MIMIC-IV-ECG and the external datasets PTB-XL and HEEDB, CARP has the lowest error on 21/24 clinical measurements. Ablations show that the world model pretraining is essential: without it, real-to-synthetic AUPRC falls 55.2%. Furthermore, while morphology supervision lowers all nine measurement errors and nearly halves FID (6.6 to 3.5). Real-to-synthetic macro-AUPRC on CARP’s ECGs reaches 80.3%, 24.0 points above the best prior generator, and augmenting rare-condition classifiers with CARP ECGs improves AUPRC over real-only training by 10.7% in-domain and by 20.1% and 15.3% on the out-of-domain datasets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
electrocardiogram, cardiology, medical AI, clinical AI, AI for healthcare

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_TKRAOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents CARP, a framework for text-conditioned ECG generation that combines a pretrained world model with a rectified-flow generator and an auxiliary clinical measurement loss. The proposed objective uses soft probabilities for ECG wave components to estimate their boundaries and amplitudes, enabling the computation of differentiable clinical measurements. The method also includes a lead-consistency loss to preserve the identities of individual ECG leads. Experiments on MIMIC-IV-ECG, PTB-XL, and HEEDB show improvements in clinical measurement errors, report alignment, and the utility of synthetic ECGs for training diagnostic classifiers.

Critical strengths

S1. The paper focuses on an important limitation that existing ECG generation methods cannot faithfully reproduce the morphological characteristics.

S2. The differentiable clinical measurement loss is rational and sound.

S3. The evaluation covers multiple aspects of generation quality.

Critical weaknesses

W1. The clinical measurement loss may not adequately enforce sample-specific alignment between clinical reports and generated ECGs. The proposed loss constructs soft histograms of clinical measurements and compares their cumulative distributions between real and generated ECGs. However, matching measurement distributions does not necessarily guarantee that each generated ECG exhibits the measurements appropriate for its own conditioning report. A model could potentially achieve good aggregate measurement statistics while producing individual ECGs with clinically inconsistent features.

W2. The method relies on a pretrained ECG segmentation model whose errors may affect the proposed supervision. The clinical measurement loss depends on differentiable wave delineation using SemiSegECG. Consequently, errors in estimating P-wave, QRS-complex, and T-wave boundaries can propagate into the computed measurements and the gradients used to train CARP. Such errors may be particularly consequential for abnormal ECGs, where wave morphology can deviate substantially from common patterns.

W3. The generated ECGs lack the direct assessment of the clinical validity. The paper reports substantial improvements in downstream classification performance, including for rare conditions and external datasets. These are encouraging results, but direct assessment such as cardiologist inspection would further strengthen the claims.

Questions for authors

Could the authors clarify whether the VAE used in CARP’s generative pipeline is trained from scratch or adopted from DiffuSETS? In Section 2.1, the authors state that they train a VAE to encode and decode ECG segments. However, Appendix C.3 states that all ECGs are encoded using the VAE released by DiffuSETS, and Table 7 further examines the choice of VAE encoder.

6 comments

W1 & W3

W1: The clinical measurement loss may not adequately enforce sample-specific alignment between clinical reports and generated ECGs [ ... ] A model could potentially achieve good aggregate measurement statistics while producing individual ECGs with clinically inconsistent features.

W3. The generated ECGs lack the direct assessment of the clinical validity.

We thank the reviewer for bringing up this potential possibility. From our evaluation, we see that when CARP is trained with the clinical measurement loss, it performs better along measurement error, real-to-synthetic (TRTS) classification, and FID (Table 2). Stronger results along classification and sample diversity (FID) provide evidence that CARP's objectives result in more clinically consistent features, as opposed to only improving on test measurements.

We also conducted a short cardiologist study with two board-certified cardiologists at a large academic hospital. Each cardiologist was shown generated ECGs and asked to rank CARP, DiffuSETS and SE-Diff. To prevent selection bias, we randomly sampled ECGs using the same training seed as in Appendix Table 3. The reviewer can inspect the images here. For each report, the cardiologists saw the three generated ECGs anonymized and in shuffled order, and ranked them from most to least clinically correct. We showed n = 13 ECGs across labels in RQ1 [1]. Ties were allowed. Neither cardiologist is an author or aware of this paper’s submission.

At the time of writing, one cardiologist has completed the study, with one pending. We provide the preliminary results below for cardiologist #1, and will update them once cardiologist #2 has completed the study.

ModelFirst places (of 13)Mean rank
CARP111.27
DiffuSETS32.35
SE-Diff22.38

These preliminary results show that cardiologists agree with the quantitative evaluation of CARP, showing its broad outperformance (mean rank 1.27 v/s 2.35 for DiffuSETS and 2.38 for SE-Diff).

[1] The labels were: atrial fibrillation, left bundle branch block, right bundle branch block, left axis deviation with sinus bradycardia, prolonged PR interval (first-degree AV block), sinus tachycardia with frequent PVCs, LVH with secondary repolarization abnormality, old inferior infarct, anterolateral ischemia, WPW pattern with sinus tachycardia, right ventricular hypertrophy, extensive ST elevation suggesting pericarditis, and dextrocardia.

W2

W2: The method relies on a pretrained ECG segmentation model whose errors may affect the proposed supervision.

Thank you for this pertinent concern. Empirically, we find that SemiSegECG’s segmentation is not an issue for learning, since CARP produces more accurate morphology for normal and abnormal ECGs for both RQ1 (train on synthetic, test on real), and RQ2 (train on real, test on synthetic), including, as shown in RQ1, on rare-conditions which have highly abnormal ECG morphology.

In addition, we evaluate SemiSegECG’s performance on abnormal ECGs, which is a SOTA model for ECG segmentation across morphologies. Since our existing datasets lack annotated segmentation labels, we use the LUDB dataset [1], which has 200 ECGs with cardiologist-annotated segmentation. We compare the precision and recall of detecting P, QRS and T waves within a 100ms region. We find that SemiSegECG and ECGDeli are generally well-performing tools, with SemiSegECG performing better in a number of instances.

Wave / label familySemiSegECG recallSemiSegECG precisionECGDeli recallECGDeli precision
P peak
Normal / sinus rhythm0.990.990.980.98
Sinus bradycardia0.991.000.980.99
Bundle-branch / IV block0.990.940.980.78
AV / fascicular / axis0.991.000.980.82
Atrial / ventricular ectopy0.970.970.980.89
Hypertrophy / enlargement0.990.970.980.87
Myocardial infarction1.000.991.000.81
Ischemia / ST-T0.980.990.980.88
Misc. rare1.001.001.000.38
All0.990.980.980.87
QRS peak
Normal / sinus rhythm1.001.000.991.00
Sinus bradycardia1.001.001.001.00
AF / flutter1.001.000.971.00
Bundle-branch / IV block1.001.001.001.00
AV / fascicular / axis1.001.000.991.00
Atrial / ventricular ectopy1.001.000.931.00
Hypertrophy / enlargement1.001.000.991.00
Myocardial infarction1.001.001.001.00
Ischemia / ST-T1.001.000.991.00
Misc. rare1.001.000.961.00
All1.001.000.991.00
T peak
Normal / sinus rhythm0.970.970.960.97
Sinus bradycardia0.980.980.940.94
AF / flutter0.910.960.840.86
Bundle-branch / IV block0.960.970.950.96
AV / fascicular / axis0.950.970.920.93
Atrial / ventricular ectopy0.970.970.890.97
Hypertrophy / enlargement0.960.970.930.95
Myocardial infarction0.960.960.910.91
Ischemia / ST-T0.950.960.930.93
Misc. rare0.920.970.810.85
All0.960.970.950.96

[1] Yusipov, I. et al (2021). Lobachevsky University Electrocardiography Database. https://www.physionet.org/content/ludb/1.0.1/

Q1

Q1: Could the authors clarify whether the VAE used in CARP’s generative pipeline is trained from scratch or adopted from DiffuSETS?

We clarify that CARP's generative VAE is trained from scratch and is different from the DiffuSETS VAE used in our evaluation. CARP's VAE (7M parameters) encodes each ECG into a sequence of latent tokens. In contrast, the DiffuSETS VAE (30M parameters) encodes an entire ECG into a single latent. This architectural difference is motivated by the original Cosmos 3's tokenization scheme, which represents inputs such as videos as sequences of latent tokens.

For downstream classification evaluation, we used the DiffuSETS VAE to ensure fair comparability with prior work. This evaluation choice is independent of CARP's generative architecture, since the DiffuSETS VAE encodes the synthetic ECGs after CARP has produced them. We additionally ablate the choice of classifier input representation in Table 7, comparing DiffuSETS VAE embeddings against raw ECG waveforms, and find that our conclusions remain consistent.

Reviewer_6xcXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CARP adapts NVIDIA Cosmos 3, a pretrained multimodal world model, to 10-second 12-lead text-to-ECG generation via rectified flow in a newly trained ECG-VAE latent space. Its second component is a differentiable measurement loss: a soft-segmentation network (SemiSegECG) turns P/QRS/T timing and amplitude measurements into differentiable histograms whose cumulative distributions for the generated ECG are matched to those of its paired real target. On MIMIC-IV-ECG and external PTB-XL/HEEDB, CARP reports the lowest error on 21/24 clinical measurements, real-to-synthetic macro-AUPRC of 80.3% (vs 56.3% DiffuSETS, 41.0% SE-Diff), and large gains from augmenting rare-condition classifiers (+10.7 points in-domain, +20.1/+15.3 out-of-domain). Ablations (8-epoch models) are presented as showing both world-model pretraining and the measurement loss are necessary.

Critical strengths

(1) The downstream-utility evidence is strong and evaluated as a practitioner would use it: synthetic augmentation across 14 label families and 5 rare conditions, in-domain and on two external datasets without retraining, against real-data, classic-augmentation and 200-real-positive oracle references — CARP is the only generator that helps out-of-domain on rare labels and approaches the oracle. (2) The soft-segmentation measurement objective is well motivated and technically concrete (soft onsets/offsets, cross-beat-penalised measurement histograms), reusable beyond ECG, with diagnostically specific qualitative cases and a clever implied-feature (axis) probe.

Critical weaknesses

(1) The "world-model prior is essential" claim is confounded: the only control random-initialises the same 4B architecture and trains it for 8 epochs (vs 32), where it underperforms much smaller from-scratch baselines — this shows the architecture is hard to optimise from scratch at this budget, not that world-model pretraining specifically transfers. No other pretrained backbone, no compute-matched or trained-to-convergence control is provided. (2) The headline measurement results are close to tautological: the metric loss directly optimises the same measurement quantities reported in Tables 1/10 against the paired real ECG; evaluating with ECGDeli instead of SemiSegECG changes the implementation of the ruler, not the quantity optimised, so "21/24 best" is not independent evidence of text alignment. The metric-loss ablation is also small/noisy and partly contradicts the abstract. (3) The SE-Diff baseline was handicapped by replacing its LLM retrieval pipeline (its morphology-label scores, e.g. LBBB 7.0%, are implausibly low), inflating part of the headline gap; the cleaner DiffuSETS comparison should be the narrative. Diversity/prototypicality is unexamined (generated ECGs align on unstated axis more often than real ECGs do; worst FID on HEEDB) and there is no memorisation/privacy analysis despite the PHI-sharing motivation.

Questions for authors

(1) Which Cosmos parameters are frozen vs fine-tuned, and what are the VAE/token dimensions? (2) At equal epochs/compute, how do a 32-epoch random-init model and a comparably sized non-world-model pretrained backbone perform? (3) Exactly which reports conditioned the 190 synthetic positives per rare condition — any test-split patients? (4) Was SemiSegECG trained on PTB-XL/HEEDB, and if so in what sense are they external? (5) Can you show the gemma swap does not explain SE-Diff's morphology scores? (6) What is per-condition sample diversity and nearest-neighbour distance of generated ECGs to training ECGs?

6 comments

W1 & Q2

W1: The "world-model prior is essential" claim is confounded: the only control random-initialises the same 4B architecture and trains it for 8 epochs (vs 32)

Q2: At equal epochs/compute, how do a 32-epoch random-init model and a comparably sized non-world-model pretrained backbone perform?

Thank you for raising these important questions regarding the value of the world model prior and the convergence of CARP. To address them, we present new ablations with different prior initializations, all trained to 32 epochs as opposed to 8 originally. In the first ablation (w/ random init), we randomly initialize the generator. In the second new ablation (w/o world-model pretraining), we follow Cosmos's training protocol and initialize the generator stack with only the language stack’s pretrained weights, and then train with the same objective as CARP. This ablation isolates the benefits of Cosmos's world-model training. Additionally, we train the ablation without the clinical metric auxiliary loss (w/o L_metric) to 32 epochs. As the table below shows, these ablations also show that CARP’s world-model pretraining and loss functions are essential improvements, particularly for morphology features, out-of-domain classification, ECG measurements, and sample diversity (lower FID).

CARPw/o L_metricw/ random initw/o world-model pretraining
Real-to-synthetic AUPRC (%) ↑
All labels80.3 [78.5, 82.2]78.5 [76.5, 80.4]62.8 [60.8, 65.0]75.2 [73.4, 77.1]
Morphology77.1 [74.5, 79.8]74.8 [71.7, 77.7]58.4 [55.4, 61.7]71.1 [68.4, 74.0]
Rate/rhythm89.5 [87.1, 91.5]88.1 [85.8, 90.4]71.1 [68.2, 74.2]85.4 [83.0, 87.8]
Measurement error (ms) ↓
PR17.318.218.518.1
QRS9.210.19.89.5
QT20.321.320.820.5
Mean RR11.311.714.813.9
Other
Axis alignment, stated (%) ↑94.694.286.092.2
FID ↓2.74.74.35.1
Downstream rare-label classification, synthetic-to-real AUPRC (%) ↑
MIMIC-IV (in-domain)85.5 ± 0.485.3 ± 0.768.1 ± 2.282.2 ± 1.3
PTB-XL (out-of-domain)78.7 ± 2.475.9 ± 2.355.7 ± 0.873.1 ± 6.3
HEEDB (out-of-domain)74.4 ± 2.972.0 ± 3.054.8 ± 0.868.7 ± 1.6

We also provide convergence curves over the 32 epochs (train and val loss), demonstrating that the CARP ablations converge before 32 epochs. Each training run accounts for 192 GPU-hours on A100s (8 A100s x 24 hrs).

[1] Cosmos 3 Technical Report

W2

W2: The headline measurement results are close to tautological: the metric loss directly optimises the same measurement quantities [ ... ] so "21/24 best" is not independent evidence of text alignment.

We thank the reviewer for raising this important distinction. We note that our experiments in RQ2 comprehensively evaluate text alignment in three ways: (a) train on real, test on synthetic (TRTS), (b) measurement and (c) report alignment. For instance in the report alignment experiment, we evaluate whether CARP can follow implied rules that are not directly stated in the report or optimized via the loss function. We found that CARP ECGs contain a cardiac axis that is aligned with the text report in 93.3% of cases where the axis is implied by the text, compared to the next-best DiffuSETS at 60% (Fig. 4c).

To reinforce our findings and address the reviewer's concern re: text alignment, we introduce new experiments to evaluate whether CARP can follow implied, diagnosis-specific rules that are not included in the text nor the metric loss. We conducted the experiment as follows: for each rare condition evaluated in RQ1, we used the FDA approved GE12SL manual [1] to identify the diagnosis-specific morphology rule; for instance, “dextrocardia” had the rule “negative QRS area in Lead I”. We then computed the fraction of generated ECGs that followed the morphology rule, and included real ECGs as a reference on the MIMIC dataset. We excluded rules that do not use intervals or amplitudes (e.g. pericarditis has a “saddle-shaped wave” rule), and some rules are disjunctive (i.e. optional), which is why not all rules are at 100% real prevalence. As shown in the table below, CARP has much higher fidelity at producing diagnosis-specific morphology than SE-Diff and DiffuSETS.

ConditionDiagnosis-specific morphologyRealCARPSE-DiffDiffuSETS
WPWSlurred upstroke, QRS > 110 ms87864645
RVHR > S (V1)85703242
90° < Axis ≤ 180°3741321
Posterior MIR > S (V2)75531422
ST depression (V1–V3)5331414
PericarditisST elevation (≥3 leads)68711041
ST depression (aVR)68671743
DextrocardiaNegative QRS (I)85731428
Positive QRS (aVR)89841926

Finally, we also conducted a cardiologist study to definitively assess whether CARP's ECGs are better aligned to the report text than baselines (see reply to Reviewer TKRA, W3).

[1] GE Healthcare. Marquette 12SL ECG Analysis Program, Physician’s Guide, 2056246-002C (12SL v23), 2019. GE Healthcare guide

W3 & Q5

W3: The SE-Diff baseline was handicapped by replacing its LLM retrieval pipeline [ ... ] the cleaner DiffuSETS comparison should be the narrative.

Q5: Can you show the gemma swap does not explain SE-Diff's morphology scores?

The reviewer raises an important point about the baseline SE-Diff’s performance being low on certain morphology classes, such as LBBB, on the real-to-synthetic evaluation. We note that the original SE-Diff paper uses the gpt-4.1-mini model for report generation, which we cannot replicate since MIMIC guidelines prohibit using credentialed data for closed-source API calls with data retention [1]. To comply with the MIMIC policy, we used the open-source gemma-4-26B-A4B-it model locally.

To address the reviewer’s concerns as suggested, we ablated the choice of gemma-4-26B-A4B-it with two leading open-source models, Beam 1.0 and DeepSeek V4.1 Flash (1017), both of which are stronger than gpt-4.1-mini.

ConditionGemma-4-26B-A4B-itBeam 1.0DeepSeek V4.1 Flash
LBBB9.1 ± 3.36.4 ± 1.97.6 ± 2.9
1st-degree AV block8.3 ± 1.58.9 ± 1.711.0 ± 2.8

As shown, ablating the choice of LLM for more capable models does not affect SE-Diff’s performance significantly. We note that SE-Diff performance remains reasonable for rhythm-based labels, such as sinus bradycardia (see Figure 5), as well as across labels on RQ2 and ECG measurement tasks. We posit that the low-performance on morphology-related classes on real-to-synthetic evaluation may be failure modes of the specific method, likely due to lower sample diversity (as evidenced by higher FID of 17.54 v/s 2.72 for CARP).

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

The paper proposes CARP, a text-to-ECG generation framework that adapts a pretrained Cosmos world model through an ECG VAE and rectified-flow training. It introduces a soft-segmentation-based clinical measurement loss to improve morphological alignment with conditioning reports. Experiments on MIMIC-IV-ECG, PTB-XL, and HEEDB report improved clinical fidelity and downstream rare-condition classification.

Critical strengths
  1. The paper addresses clinically important failures in text-to-ECG generation and evaluates morphology, cardiac axis, report consistency, and downstream utility beyond pointwise reconstruction error.

  2. CARP shows substantial improvements over the evaluated generators across classification tasks and external datasets. Independent ECGDeli evaluation and the raw-waveform experiment in Table 7 strengthen the empirical evidence.

Critical weaknesses
  1. The code from anonymous repository seems using a different clinical loss, adding supervision not described in the paper, and omitting the low-noise restriction in Eq. (8). The evaluation classifier also differs from the stated architecture. These inconsistencies make it unclear whether the experiments evaluate the proposed method. The authors should provide the exact code version and configurations underlying the reported results.

  2. The pretrained vs random initialization comparison does not isolate the value of world-model priors from generic pretraining, model capacity, or optimization advantages. All ablations run for eight epochs, while the main model trains for 32. Without convergence analysis or comparable pretrained-backbone controls, the distinct contribution of world-model pretraining remains unclear.

  3. The clinical loss matches interval and amplitude distributions separately, which does not preserve how these measurements co-occur within individual beats or their temporal ordering. Matching these distributions therefore does not necessarily preserve diagnosis-defining patterns. The paper lacks targeted evaluations showing that this surrogate improves such clinical relationships beyond what the pretrained generator already captures.

  4. The MIMIC-IV-ECG reports which are used for model training are machine-generated. Without assessing report–waveform agreement, systematic interpretation errors could affect both training and evaluation, making label agreement insufficient to establish clinical correctness. Although external evaluation provides useful evidence, expert-adjudicated assessment of the rare conditions would help resolve this concern.

  5. Removing LmetricL_{\mathrm{metric}} changes macro-AUPRC from 73.0 to 71.7 and morphology AUPRC from 67.0 to 65.3, while four of seven morphology labels improve without it. Although FID improves, the loss is not ablated on the headline rare-condition augmentation task or compared with simpler clinical supervision. Its distinct clinical benefit remains insufficiently established.

Questions for authors
  1. Can you provide convergence curves and a comparable alternative pretrained generative backbone to distinguish world-model priors from generic pretraining and optimization advantages?

  2. How accurately do the proposed soft measurements approximate annotated measurements across different segmentation confidence levels, missing waves, and pathological morphologies?

  3. Does LmetricL_{\mathrm{metric}} improve rare-condition augmentation after the full 32-epoch schedule? How does it compare with simpler clinical-feature supervision, with uncertainty reported for paired performance differences?

  4. How many target-condition examples are available during generator training, and where do the 190 synthetic conditioning reports originate? Can you compare against baselines with matched access to this supervision?

  5. Beyond the aggregate measurements in Tables 1 and 10, can you quantify diagnosis-specific morphology across each condition’s evaluation cohort, such as QRS shape, R/S amplitude patterns, and T-wave polarity where relevant?

9 comments

W1

W1: The code from anonymous repository seems using a different clinical loss [ ... ] The authors should provide the exact code version and configurations underlying the reported results.

We appreciate the reviewer’s detailed attention to our paper and implementation. We note that, for transparency into our research process, multiple earlier versions of our loss function were included in the repo, as well as other auxiliary loss candidates which were not ultimately used. Furthermore, we can confirm that the stated architectures for classification are correct, for which we note there were two types (VAE for encoding, ResNet for multi-label classification), all of which follow SE-Diff.

We apologize for these points of confusion and have updated the repository accordingly to include a reproduction script with precise config files (ssg_simple_r20_w025_no_qrs_corr_e32.yaml) and training script (rebuttals/reproduction/train_carp.sh).

W3

The clinical loss matches interval and amplitude distributions separately [ ... ] The paper lacks targeted evaluations showing that this surrogate improves such clinical relationships beyond what the pretrained generator already captures.

We thank the reviewer for pointing out this possibility. We note that RQ3 already explicitly evaluates whether the auxiliary loss improves modeling of diagnosis-specific morphology beyond what the standard rectified-flow objective captures. In Table 12, removing the clinical loss degrades FID (3.5 to 6.6) and reduces morphology real-to-synthetic AUPRC (67.0% to 65.3%). Furthermore, Table 2 shows that the clinical loss improves axis alignment from 88.8% to 91.7%. These results directly support that the clinical measurement loss improves representation of diagnosis-specific morphology. We also note that, although the clinical metric distributions are measured separately, they are optimized jointly with the rectified flow objective, allowing for covariance and temporal ordering to be modeled. Lastly, our train-on-real, test-on-synth (TRTS) results in RQ2 further illustrate that CARP captures diagnosis-specific morphology well (macro-AUPRC of 80.3%), significantly outperforming DiffuSETS (56.3%) and SE-Diff (41.0%), both of which also use a standard reconstruction objective.

W4

W4: MIMIC-IV-ECG reports which are used for model training are machine-generated [ ... ] making label agreement insufficient to establish clinical correctness [ ... ] Although external evaluation provides useful evidence, expert-adjudicated assessment of the rare conditions would help resolve this concern.

We thank the reviewer for this important question. We note that the labels for PTB-XL are already fully expert-adjudicated by consensus from two board-certified cardiologists, including rare conditions, and our sampling is detailed in Section C.5. The experimental results in RQ1 and RQ2 demonstrate that CARP far outperforms baseline methods on PTB-XL, regardless of any bias inherent to MIMIC-IV-ECG. For instance, on PTB-XL, classifiers augmented with CARP ECGs reach 78.7% AUPRC on rare conditions, against 56.5% for DiffuSETS and 49.8% for SE-Diff (RQ1).

Furthermore, the MIMIC-IV-ECG reports are produced from different FDA-approved providers (Philips, GE, and Spacelabs), which mitigates single-provider bias. The out-of-domain evaluation extensively evaluates transfer to datasets from different geographies (Germany, North America) and interpretation algorithms (Phillips, GE12SL). Finally, in order to ensure a fair comparison, CARP and the baselines (SE-Diff, DiffuSETS) use the same MIMIC-IV-ECG dataset for training.

Finally, to directly assess clinical correctness, as the reviewer suggests, we conducted a short cardiologist evaluation with two board-certified cardiologists at a large academic hospital. Each cardiologist was shown generated ECGs and asked to rank CARP, DiffuSETS and SE-Diff. To prevent selection bias, we randomly sampled ECGs using the same training seed from Appendix Table 3. For each report, the cardiologist saw the three generated ECGs anonymized and in shuffled order, and ranked them from most to least clinically correct. We showed n = 13 ECGs across labels in RQ1 [1]. Ties were allowed. Neither cardiologist is an author or aware of this paper’s submission.

At the time of writing, one cardiologist has completed the study, with one pending. We provide the preliminary results below for cardiologist #1, and will update them once cardiologist #2 has completed the study.

ModelFirst places (of 13)Mean rank
CARP111.27
DiffuSETS32.35
SE-Diff22.38

These preliminary results show that cardiologists agree with the quantitative evaluation of CARP, showing its broad outperformance (mean rank 1.27 v/s 2.35 for DiffuSETS and 2.38 for SE-Diff).

[1] The labels were: atrial fibrillation, left bundle branch block, right bundle branch block, left axis deviation with sinus bradycardia, prolonged PR interval (first-degree AV block), sinus tachycardia with frequent PVCs, LVH with secondary repolarization abnormality, old inferior infarct, anterolateral ischemia, WPW pattern with sinus tachycardia, right ventricular hypertrophy, extensive ST elevation suggesting pericarditis, and dextrocardia.

W2 & Q1

W2: The pretrained vs random initialization comparison does not isolate the value of world-model priors from generic pretraining

Q1: Can you provide convergence curves and a comparable alternative pretrained generative backbone to distinguish world-model priors from generic pretraining and optimization advantages?

Thank you for raising these important questions regarding the value of the world model prior and the convergence of CARP. To address them, we present new ablations with different prior initializations, all trained to 32 epochs as opposed to 8 originally. In the first ablation (w/ random init), we randomly initialize the generator. In the second new ablation (w/o world-model pretraining), we follow Cosmos's training protocol and initialize the generator stack with only the language stack’s pretrained weights, and then train with the same objective as CARP. This ablation isolates the benefits of Cosmos's world-model training. Additionally, we train the ablation without the clinical metric auxiliary loss (w/o L_metric) to 32 epochs. As the table below shows, these ablations also show that CARP’s world-model pretraining and loss functions are essential improvements, particularly for morphology features, out-of-domain classification, ECG measurements, and sample diversity (lower FID).

CARPw/o L_metricw/ random initw/o world-model pretraining
Real-to-synthetic AUPRC (%) ↑
All labels80.3 [78.5, 82.2]78.5 [76.5, 80.4]62.8 [60.8, 65.0]75.2 [73.4, 77.1]
Morphology77.1 [74.5, 79.8]74.8 [71.7, 77.7]58.4 [55.4, 61.7]71.1 [68.4, 74.0]
Rate/rhythm89.5 [87.1, 91.5]88.1 [85.8, 90.4]71.1 [68.2, 74.2]85.4 [83.0, 87.8]
Measurement error (ms) ↓
PR17.318.218.518.1
QRS9.210.19.89.5
QT20.321.320.820.5
Mean RR11.311.714.813.9
Other
Axis alignment, stated (%) ↑94.694.286.092.2
FID ↓2.74.74.35.1
Downstream rare-label classification, synthetic-to-real AUPRC (%) ↑
MIMIC-IV (in-domain)85.5 ± 0.485.3 ± 0.768.1 ± 2.282.2 ± 1.3
PTB-XL (out-of-domain)78.7 ± 2.475.9 ± 2.355.7 ± 0.873.1 ± 6.3
HEEDB (out-of-domain)74.4 ± 2.972.0 ± 3.054.8 ± 0.868.7 ± 1.6

We also provide convergence curves over the 32 epochs (train and val loss), demonstrating that the CARP ablations converge before 32 epochs. Each training run accounts for 192 GPU-hours on A100s (8 A100s x 24 hrs).

[1] https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf

Reviewer_gg4YOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes CARP, a text-conditional ECG generation framework that transfers a pre-trained world model and incorporates differentiable clinical morphological measurement supervision to improve the alignment between synthetic ECGs and diagnostic reports. The authors trained the model on approximately 556,000 MIMIC-IV-ECG records and validated it on two external datasets. Experiments demonstrate that CARP outperforms existing generative methods in clinical measurements, report consistency, and data augmentation for rare diseases. Its core contribution is integrating pre-trained generative capabilities with ECG-specific clinical constraints, such that improvements in generation quality translate into tangible benefits for real-world downstream tasks.

Critical strengths
  1. A targeted improvement in clinical alignment.
    The paper addresses the practical problem of generating ECGs that appear plausible but lack the features required to support the specified diagnosis. Its differentiable clinical-measurement supervision directly constrains morphological details. Compared with improving distribution-level metrics alone, this design more closely addresses the core requirements of text-to-ECG generation.

  2. Evaluation covers both generation quality and practical utility.
    The experiments assess clinical measurements, report alignment, synthetic-data augmentation, and cross-dataset generalization, providing a reasonably comprehensive body of evidence. Using SemiSegECG for training supervision and ECGDeli for measurement-based evaluation also reduces potential bias from relying on the same measurement tool for both.

Critical weaknesses
  1. The distinctive value of world-model pretraining is not fully isolated.
    The randomly initialized generator ablation is trained for only eight epochs (Table 2), whereas the main training runs for 32 epochs. This experiment supports the advantage of pretraining under a matched short-term training budget, but does not establish whether training from scratch could eventually close the gap. It also does not distinguish the benefits of pretrained initialization in general from those of priors specific to a world model.

  2. The mathematical description of the core clinical loss requires clarification.
    The window in Equation (6) depends on hard labels with stopped gradients, so the claim that every non-differentiable step has been replaced needs qualification. A smoothing function is defined immediately before Equation (7) but is not used in the equation, and the normalization of measurements with different units is insufficiently specified.

Questions for authors
  1. Contribution of pretraining and the training budget for ablations.
    All variants in Table 2 are trained for eight epochs. Has the randomly initialized generator sufficiently converged, and does the performance gap persist with longer training? Furthermore, could the authors clarify whether the current evidence supports the benefits of pretrained initialization in general or a distinctive advantage of world-model priors over other forms of pretraining?

  2. Implementation and differentiability of the clinical-measurement loss.
    Equation (6) uses detached argmax labels to determine the window. Does this mean that gradients cannot propagate through the window-selection operation? Does the implementation of Equation (7) use the absolute value or the previously defined smoothing function ρ\rho? Additionally, how are measurements with different units, such as durations and amplitudes, normalized, and which specific term receives the rhythm weight of 20 listed in Appendix Table 3?

  3. Scope of the low-resource augmentation experiments.
    The downstream classifiers use only ten real positive samples, but the generator is trained on the full MIMIC-IV-ECG training set and may have encountered substantially more examples of the same conditions. Could the authors report the number of positive examples available to the generator for each condition and distinguish between a setting with limited downstream labels and one in which only ten positive samples are available to the entire system? Could they also provide evaluation results at more realistic disease prevalences to complement the AUPRC results obtained on balanced test sets?

4 comments

W1 & Q1

W1: The distinctive value of world-model pretraining is not fully isolated.

Q1: Contribution of pretraining and the training budget for ablations.

Thank you for raising these important questions regarding the value of the world model prior and the convergence of CARP. To address them, we present new ablations with different prior initializations, all trained to 32 epochs as opposed to 8 originally. In the first ablation (w/ random init), we randomly initialize the generator. In the second new ablation (w/o world-model pretraining), we follow Cosmos's training protocol and initialize the generator stack with only the language stack’s pretrained weights, and then train with the same objective as CARP. This ablation isolates the benefits of Cosmos's world-model training. Additionally, we train the ablation without the clinical metric auxiliary loss (w/o L_metric) to 32 epochs. As the table below shows, these ablations also show that CARP’s world-model pretraining and loss functions are essential improvements, particularly for morphology features, out-of-domain classification, ECG measurements, and sample diversity (lower FID).

CARPw/o L_metricw/ random initw/o world-model pretraining
Real-to-synthetic AUPRC (%) ↑
All labels80.3 [78.5, 82.2]78.5 [76.5, 80.4]62.8 [60.8, 65.0]75.2 [73.4, 77.1]
Morphology77.1 [74.5, 79.8]74.8 [71.7, 77.7]58.4 [55.4, 61.7]71.1 [68.4, 74.0]
Rate/rhythm89.5 [87.1, 91.5]88.1 [85.8, 90.4]71.1 [68.2, 74.2]85.4 [83.0, 87.8]
Measurement error (ms) ↓
PR17.318.218.518.1
QRS9.210.19.89.5
QT20.321.320.820.5
Mean RR11.311.714.813.9
Other
Axis alignment, stated (%) ↑94.694.286.092.2
FID ↓2.74.74.35.1
Downstream rare-label classification, synthetic-to-real AUPRC (%) ↑
MIMIC-IV (in-domain)85.5 ± 0.485.3 ± 0.768.1 ± 2.282.2 ± 1.3
PTB-XL (out-of-domain)78.7 ± 2.475.9 ± 2.355.7 ± 0.873.1 ± 6.3
HEEDB (out-of-domain)74.4 ± 2.972.0 ± 3.054.8 ± 0.868.7 ± 1.6

We also provide convergence curves over the 32 epochs (train and val loss), demonstrating that the CARP ablations converge before 32 epochs. Each training run accounts for 192 GPU-hours on A100s (8 A100s x 24 hrs).

[1] Cosmos 3 Technical Report

W2 & Q2(a)

W2: The mathematical description of the core clinical loss requires clarification.

Q2(a) Equation (6) uses detached argmax labels to determine the window. Does this mean that gradients cannot propagate through the window-selection operation?

We would be happy to clarify. For any wave (QRS, P, or T), the argmax only selects the end of the wave (i.e. the “offset”). The gradients for the start of the wave (“onset”) and the voltages within the wave are propagated fully. This allows the generative model to learn the amplitude (i.e. the peak voltage) of the wave. The offset itself still gets gradients from the wave durations (e.g., QRS duration) in Eqs. (4)–(5).

W2 & Q2(b)

Q2(b) Does the implementation of Equation (7) use the absolute value or the previously defined smoothing function (\rho)?

Thank you for pointing out this typo, we omitted the rho in Equation 7 and have corrected our manuscript. The correct text should be:

ℓm,l=Δm∑bρ(F^m,l(b)−Fm,l(b))\ell_{m,l} = \Delta_m \sum_b \rho\left( \hat F_{m,l}(b)-F_{m,l}(b) \right)

W2 & Q2(c)

Additionally, how are measurements with different units, such as durations and amplitudes, normalized, and which specific term receives the rhythm weight of 20 listed in Appendix Table 3?

We normalize each measurement histogram to sum to 1, which ensures that the histograms are invariant to heart rate. We then compute the distance between the generated and real histograms, which gives the error in seconds for durations and in mV for amplitudes. We average these errors over all measurements with equal weight. Finally, the rhythm weight refers to the RR-interval loss.

Q3

Scope of the low-resource augmentation experiments [ ... ] Could they also provide evaluation results at more realistic disease prevalences to complement the AUPRC results obtained on balanced test sets?

Thank you for this question. All generative models (CARP, SE-Diff and DiffuSETS) were trained on several examples of each condition, including rare labels. Appendix Table 5 provides the counts for the full MIMIC set, we used 70% of it for training, as specified in Appendix C.1. For instance, the MIMIC dataset has 306 ECGs with dextrocardia (of which 70% are used for training, i.e. 214 ECGs), at 0.039% prevalence.

To clarify: in the downstream classifier evaluation, we deliberately limit to k samples following prior work (SE-Diff) to study the transfer to low-resource settings, where for instance plenty of labels from North America or Europe are available, but few or none from low-income countries. In such settings, a classifier trained on North American data is tested with no adaptation on data from low-resource settings, often with poor generalizability [1, 2].

To directly address the reviewer’s concern, we conduct a new experiment to evaluate performance at natural disease prevalence. We train a multi-label classifier (CNN) on only synthetic ECGs (n = 556k), and test it on real ECGs from MIMIC at their actual disease prevalence. As the reviewer requests, this new experiment differs from RQ1 which uses binary classifiers (one per label), 10 real samples, and a balanced test dataset. We also provide the real ECG classifier performance as a reference upper bound. In the table below, we show the AUROC results for the rare conditions. We find that, like in RQ1, CARP outperforms baselines, and approaches the performance of a reference classifier, particularly for Pericarditis.

Label (prevalence)RealCARPDiffuSETSSE-Diff
WPW (0.06%)82.3 ± 4.474.1 ± 0.347.4 ± 1.349.5 ± 5.6
RVH (0.60%)94.8 ± 0.192.4 ± 0.682.9 ± 1.166.6 ± 1.2
PMI (0.08%)94.1 ± 1.788.3 ± 2.774.7 ± 5.768.3 ± 2.7
PCT (0.31%)97.7 ± 0.395.7 ± 0.891.4 ± 1.684.9 ± 1.4
DCT (0.04%)97.4 ± 0.795.1 ± 1.667.8 ± 5.863.2 ± 4.1

[1] Yagi et al, 2022. Cited in paper

[2] Perez et al, 2020. Cited in paper

We appreciate the reviewer's constructive feedback. In response, we conducted new pretraining ablations, clarified our measurement loss, and introduced a new evaluation showing CARP outperforming both baselines across five rare conditions even at natural disease prevalence. We also conducted a cardiologist study which showed CARP's superior text alignment. We hope these address the reviewer's concerns and we would welcome further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.