acceptodds
Under review as a conference paper at ICLR 2027

Physiology-Guided World Models for Intervention-Aware Cardiac Simulation

Abstract

Electrocardiogram (ECG)-based models have achieved strong diagnostic performance, yet remain limited in modeling how cardiac dynamics evolve under external interventions. Existing approaches primarily focus on static prediction, with limited capacity to capture ECG changes under different pharmacological conditions. In this work, we propose an ECG World Model for action-conditioned predictive simulation of cardiac electrophysiology. Our framework integrates physiological ordinary differential equation (ODE) priors into latent diffusion dynamics through energy regularization. This structural guidance encourages the generation of physiologically plausible post-intervention ECG trajectories while reducing generative inconsistencies. Building on these simulations, we introduce an uncertainty-aware evaluation strategy that leverages stochastic diffusion sampling to estimate both the expected risk score and its variability, supporting comparative assessment of candidate interventions. We evaluate our method across diverse settings, including controlled drug-response datasets and real-world clinical records. Beyond improvements in waveform prediction, experimental results demonstrate better preservation of clinically relevant ECG intervals and more accurate prediction of the magnitude and direction of drug-induced risk changes. These results highlight the potential of our approach for intervention-aware cardiac simulation and simulation-based assessment of candidate treatments.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
World Model, ECG

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Mo9pOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an ECG world model for predicting drug-induced changes in ECG waveforms. It incorporates physiological ODE priors into a latent diffusion model through energy regularization. A frozen ECG foundation model scores the generated ECGs, and multiple stochastic samples are used to estimate the mean and variability of diagnostic risk scores for comparing candidate interventions. Evaluations on controlled drug-response datasets and real-world clinical records report improvements in waveform prediction, preservation of clinically relevant ECG intervals, and prediction of drug-induced changes in diagnostic scores.

Critical strengths

Broad experimental coverage. The authors evaluated the method using both controlled drug-response datasets and real-world clinical data, complemented by cross-dataset testing, ablation studies, and robustness analyses. The evaluation encompassed waveform errors, ECG intervals, and changes in diagnostic scores, providing empirical support for the method's effectiveness from multiple perspectives.

Critical weaknesses

(1). Insufficient validation of risk scores and uncertainty components. The authors construct a risk proxy by averaging probabilities across multiple heterogeneous diagnostic labels but fail to adequately justify the metric's applicability when comparing different medications. Furthermore, they do not directly demonstrate that incorporating standard deviation—as opposed to relying solely on the mean—leads to superior intervention selection. The experimental results regarding "diagnostic changes" primarily reflect the fidelity of the generation process rather than the effectiveness of the ranking module; this undermines the persuasiveness of the paper's core contribution: the comparative evaluation of interventions.

(2). The applicability of physiological priors to complex heart rhythms remains unclear. As the authors acknowledge in the discussion of limitations, the framework's reliance on classical ODE priors may struggle to accommodate severely irregular rhythms, such as ventricular fibrillation. The paper does not sufficiently explain how the generated ECG and subsequent diagnostic scores are affected when the prior is mismatched. Given that physiological constraints are central to the method's design, this uncertainty limits my ability to assess its reliability in complex pathological scenarios.

Questions for authors
  1. Ventricular fibrillation is a clinically significant factor when evaluating pharmacological interventions. The paper acknowledges the limitations of its ODE prior in such scenarios. Could the authors explain how the model handles these situations and discuss the impact on the generated ECGs and subsequent risk scores?

  2. The paper states that the framework supports the representation of continuous dosages, yet discrete intervention tokens were used in the experiments. What experimental evidence supports the model's ability to handle continuous dosages?

  3. Could the authors clarify the specific meaning of the values ​​in the predictive distribution and explain their relationship to clinical risk?

4 comments

W1 and Q3. Interpretation and validation of the predictive score distribution

  • We compare generated and observed ECGs using the mean probability across the frozen classifier's 17 labels. These labels cover sinus rhythm, pathological findings, and signal quality, so a higher mean has no uniform interpretation as clinical risk. For a given input and regimen, the mean and standard deviation of this diagnostic summary across generated ECGs describe its predictive distribution. Across 200 held-out test samples, predicted and observed changes in the summary had a Pearson correlation of 0.620, a Spearman correlation of 0.598, and 75.5% directional agreement.

  • We agree that dispersion across generated ECGs should not be interpreted as calibrated clinical uncertainty. Our additional evaluation reveals substantial undercoverage of nominal 90% prediction intervals, with mixed CRPS results across endpoints. We therefore restrict our conclusions to empirical predictive-distribution assessment and exploratory uncertainty-aware QTc ranking, rather than calibrated clinical risk estimation.

    Full EPK increased coverage for latent-variable prediction intervals and yielded a higher CRPS. At 90% nominal coverage, empirical coverage increased from 69.66% to 80.09%, and CRPS increased from 0.097785 to 0.103025.

Predicted quantityCRPS (full EPK / no EPK)Empirical coverage at 90% nominal coverage (full EPK / no EPK)Prediction interval width (full EPK / no EPK)
Latent dimensions0.103025 / 0.09778580.09% / 69.66%0.61641 / 0.36055
Lead II QTc25.536 / 16.245 ms63.59% / 69.74%82.812 / 68.338 ms
Lead aVF QTc16.063 / 16.475 ms75.99% / 70.63%76.923 / 66.149 ms
  • We compared rankings based on the predictive mean alone and on the mean plus a weighted standard deviation. The cost measures positive QTc prolongation from baseline.

    λ\lambda weights the standard deviation of the predicted cost. At λ=0\lambda=0, the ranking uses the predictive mean alone. ΔQTc is the change from the pre-dose baseline, and ΔΔQTc subtracts the corresponding change under matched placebo. Regret is the difference in observed cost between the model's top-ranked regimen and the best regimen according to the observations. Top-1 measures agreement on the best regimen, and Kendall τb\tau_b measures agreement across the full ranking.

    For lead II ΔQTc without placebo adjustment, λ=1\lambda=1 gave the best results across all three ranking metrics. Relative to the mean-only ranking, regret fell from 10.300 to 8.812 ms and Top-1 agreement increased from 55.62% to 58.81%.

EndpointλRegret (ms) ↓Top-1 (%) ↑Kendall τb ↑
II ΔQTc010.30055.620.1426
II ΔQTc0.59.24955.400.1426
II ΔQTc18.81258.810.1815
II ΔQTc29.60357.600.1774
aVF ΔQTc010.16056.160.2642
aVF ΔQTc0.510.12055.350.2844
aVF ΔQTc110.34854.380.2726
aVF ΔQTc210.27853.900.2687
II ΔΔQTc010.62656.670.2690
II ΔΔQTc0.510.92856.730.2654
II ΔΔQTc110.83057.840.2636
II ΔΔQTc210.83057.840.2675
aVF ΔΔQTc012.13950.180.2621
aVF ΔΔQTc0.511.58650.920.2599
aVF ΔΔQTc111.70950.120.2542
aVF ΔΔQTc212.09550.120.2477

W2 and Q1. Irregular rhythms and prior mismatch

  • The MIMIC test set includes ECGs with irregular rhythms. We report MSE over the entire test set. The synthetic jitter and latent-space Gaussian noise experiments in Appendix B.9 measure the dispersion of generated samples under perturbations. Waveform prediction accuracy and the reliability of diagnostic outputs in ventricular fibrillation remain to be validated. Accordingly, we explicitly scope our clinical claims to the evaluated sinus and non-arrhythmic/mildly irregular regimes, and treat extreme arrhythmia settings such as ventricular fibrillation as an out-of-distribution regime requiring dedicated validation before clinical deployment.

  • Full EPK reduced the dispersion of generated samples under synthetic perturbations. Figure 9 in Appendix B.9 compares full EPK, EPK at one-quarter strength, and EPK disabled. Under strong jitter, full EPK reduced the mean pairwise ℓ₂ distance from approximately 6.9 to 5.7 and the latent standard deviation from approximately 0.21 to 0.17 relative to disabling EPK. With Gaussian noise in latent space, the corresponding reductions were approximately 6.6 to 5.1 and 0.20 to 0.15. The weaker prior produced intermediate results.

  • In a separate controlled drug-response evaluation, full EPK reduced MAE for all six ECG interval endpoints. The heart rate used as a conditioning input at inference is extracted directly from the original pre-dose ECG recorded at the corresponding visit. Measured post-dose heart rate is not used as a model input.
Lead and intervalFull EPK MAE (ms) ↓No EPK MAE (ms) ↓
Lead II QTc27.82231.793
Lead II PR16.70717.945
Lead II TpTe10.41510.662
Lead aVF QTc27.90231.125
Lead aVF PR14.70516.541
Lead aVF TpTe10.14011.223

Q2. Continuous dosing and the evaluated action space

  • We evaluated 8 fixed drug regimens comprising 4 monotherapies, 3 combination therapies, and placebo. Each regimen was encoded as a separate discrete action (Appendix E.7). We have not evaluated prediction at continuous dose levels, so our empirical conclusions are limited to these fixed regimens.

Thank you for the response and additional experiments, which partially address my concerns. As the authors acknowledge, dispersion across generated ECGs should not be interpreted as calibrated clinical uncertainty. The current results support the method as a drug-conditioned ECG generative model. However, I remain unconvinced that it can reliably compare the clinical consequences of alternative interventions, a capability central to the paper’s broader “world model” claim. I will therefore maintain my current score.

We thank the reviewer for the constructive follow-up and for acknowledging that our results support ECG-WM as a drug-conditioned ECG generative model.

We would like to clarify that the primary objective of ECG-WM is to learn action-conditioned cardiac dynamics and simulate ECG responses under different pharmacological interventions. In this context, the term "world model" refers to a learned action-conditioned transition model that predicts future physiological states given the current ECG and a specified intervention. This modelling objective is distinct from developing a clinically validated treatment-selection system.

Beyond ECG generation, our randomised crossover evaluation provides quantitative evidence for comparing drug-associated electrophysiological responses across interventions. Specifically, the model demonstrates agreement with observed within-subject QTc response rankings across treatment periods, and incorporating predictive uncertainty improves selected ranking metrics. We believe these results provide meaningful initial evidence for intervention-aware simulation, while recognising that QTc-based comparisons do not capture the full range of clinical consequences associated with treatment decisions.

To make this distinction explicit, we will clarify in the revised manuscript's Limitations section that the current evaluation focuses on electrophysiological responses, and that comprehensive clinical treatment-selection applications require broader outcome assessment and further clinical validation. This clarification defines the intended scope of ECG-WM without diminishing its demonstrated contributions to physiology-guided, drug-conditioned ECG simulation and intervention-response modelling.

We appreciate the reviewer's recognition of our generative modelling results and hope this clarification helps distinguish the methodological contribution of ECG-WM from its potential future clinical applications.

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

The paper proposes a drug-conditioned latent diffusion model for predicting post-dose 12-lead ECGs from pre-dose recordings and clinical context. A cardiac ODE supplies physiological conditioning and a latent regularization target. A frozen ECG diagnostic model then scores generated samples to support candidate intervention comparison. Experiments evaluate waveform prediction, interval fidelity, diagnostic-score changes, and short autoregressive rollouts using controlled drug-response cohorts and MIMIC-derived clinical data.

Critical strengths
  1. The framework addresses a meaningful extension of ECG modeling by predicting medication-associated waveform changes from a patient’s current recording.

  2. Combining intervention conditioning with an ODE-derived temporal prior provides a concrete approach to incorporating physiological structure into multilead ECG generation.

  3. The evaluation considers clinically relevant intervals and diagnostic changes alongside waveform errors. And the clinical and cross-cohort experiments provide additional encouraging evidence of predictive utility.

Critical weaknesses
  1. The score used to compare interventions lacks a consistent clinical direction. Appendix B.10 averages 17 diagnostic probabilities, including sinus rhythm, poor data quality, and pathological findings. Higher confidence in sinus rhythm therefore increases the quantity being minimized, all else equal. Although the paper acknowledges that this is a proxy, agreement with changes in this aggregate does not establish meaningful treatment ranking. The reported evaluations concern the recorded medication and do not demonstrate that selecting an alternative intervention, or penalizing sampling variability, improves the assessment.

  2. The experiments do not convincingly isolate the benefit of correct drug-specific physiological guidance. With EPK conditioning retained, removing the energy penalty yields slightly lower signal MSE in Table 6, although other metrics improve with regularization.

  3. Pharmacological versus identity mappings produce mixed results in Table 11, and the permuted mapping is evaluated only through template statistics rather than predictive performance. These controls leave unclear whether the gains depend on pharmacologically meaningful structure or mainly on additional conditioning and generic latent regularization.

  4. The simulator supplies a drug-conditioned morphology template, but its connection to patient-specific drug responses remains unclear. Drug effects are introduced through heuristic parameter changes, without sufficiently specifying dose dependence, elapsed post-dose time, or individual sensitivity.

  5. The physiological loss measures distance to a jointly learned latent anchor, without establishing that this distance reflects physiological violations in generated ECGs. Mixed results against identity mappings and the absence of predictive evaluations with permuted mappings further weaken attribution of the gains to correct physiological knowledge.

  6. Reduced sampling variability does not establish reliability under simulator mismatch. Regularization can concentrate predictions around an incorrect prior, producing low dispersion despite substantial bias. The paper does not demonstrate that its uncertainty penalty improves intervention ranking.

Questions for authors
  1. Why are sinus rhythm and pathological findings assigned the same positive weight in the risk score? Do the intervention-ranking conclusions persist with a clinically justified score, and what evidence supports the benefit of the uncertainty penalty?

  2. Under identical architecture, training seeds, and sampling budgets, how do pharmacological, identity, and permuted mappings compare on numerical interval-change prediction? Can these comparisons separate the effects of EPK conditioning and the energy penalty?

  3. How are pre-dose and post-dose ECGs paired relative to medication administration, and is elapsed time included in the conditioning? ECGRDVQ contains multiple post-dose measurement times, making this important for defining the prediction target.

7 comments

W1 and Q1. Score direction and the value of uncertainty-aware ranking

  • The original score is a summary of diagnostic probabilities. For an ECG xx, the score is the arithmetic mean of the probabilities pj(x)p_j(x) that the frozen ECG classifier assigns to its 17 labels.

Rdiag(x)=117∑j=117pj(x).R_{\mathrm{diag}}(x)=\frac{1}{17}\sum_{j=1}^{17}p_j(x).

The labels cover sinus rhythm, pathological findings, and signal quality. An increase in any label probability raises the mean, so the direction of the score has no uniform clinical risk interpretation. We use it to compare the diagnostic probability summaries of generated and observed ECGs. Across 200 held-out test samples, the predicted and observed changes in this summary had a Pearson correlation of 0.620, a Spearman correlation of 0.598, and 75.5% directional agreement.

  • We rank candidate regimens by positive QTc prolongation. For sample ss of regimen aa, we define the cost relative to the pre-dose QTc from that regimen's own visit.

C(s)(a)=max⁡ ⁣{QTcpost,a(s)−QTcpre,a, 0}.C^{(s)}(a)=\max\!\left\{QTc^{(s)}_{\mathrm{post},a}-QTc_{\mathrm{pre},a},\,0\right\}.

We rank candidates from lowest to highest Sλ(a)S_\lambda(a), using the mean and standard deviation of the costs across the same 20 generated samples per regimen.

Sλ(a)=C‾(a)+λ SD⁡ ⁣(C(1)(a),…,C(20)(a)).S_\lambda(a)=\overline{C}(a)+\lambda\,\operatorname{SD}\!\left(C^{(1)}(a),\ldots,C^{(20)}(a)\right).

We compare λ=0,0.5,1,2\lambda=0,0.5,1,2, with λ=0\lambda=0 using the predictive mean alone. For placebo-adjusted analyses, the predicted ranking subtracts the model-predicted ΔQTc under the matched placebo action from the candidate regimen's predicted ΔQTc before taking the positive part, while the reference ranking subtracts the measured placebo ΔQTc from the measured candidate ΔQTc. We report results with and without placebo adjustment in both leads. Regret is the difference in reference cost between the model's top-ranked regimen and the best regimen according to the reference. Top-1 measures how often the top-ranked regimens agree, accounting for ties. Kendall's τb\tau_b measures agreement across the full ranking.

For lead II ΔQTc without placebo adjustment (8 participants and 164 evaluable ranking comparisons), λ=1\lambda=1 reduced regret from 10.300 to 8.812 ms relative to the mean-only ranking (paired difference −1.488 ms, 95% CI [−3.281, −0.033]). Top-1 agreement increased from 55.62% to 58.81% (paired difference +3.19 percentage points, 95% CI [0.17, 6.16]). These results are exploratory, given the limited number of independent participants and comparisons across multiple endpoints and uncertainty weights. The observed benefit of uncertainty penalisation is endpoint-dependent and does not establish superior clinical treatment selection.

EndpointλRegret (ms) ↓Top-1 (%) ↑Kendall τb ↑
II ΔQTc010.30055.620.1426
II ΔQTc0.59.24955.400.1426
II ΔQTc18.81258.810.1815
II ΔQTc29.60357.600.1774
aVF ΔQTc010.16056.160.2642
aVF ΔQTc0.510.12055.350.2844
aVF ΔQTc110.34854.380.2726
aVF ΔQTc210.27853.900.2687
II ΔΔQTc010.62656.670.2690
II ΔΔQTc0.510.92856.730.2654
II ΔΔQTc110.83057.840.2636
II ΔΔQTc210.83057.840.2675
aVF ΔΔQTc012.13950.180.2621
aVF ΔΔQTc0.511.58650.920.2599
aVF ΔΔQTc111.70950.120.2542
aVF ΔΔQTc212.09550.120.2477

W2, W3, W5, and Q2. Overall EPK performance and the role of the energy term

  • Disentangling pharmacological conditioning and physiological energy regularisation. We thank the reviewer for highlighting the need to distinguish the contributions of pharmacological conditioning and energy regularisation. Our matched five-way ablation provides evidence for two complementary, but empirically distinct, components.

    First, correct pharmacological structure matters. Full EPK achieves a waveform MSE of 0.057402, compared with 0.058772 for identity mapping and 0.059211 for permuted mapping, corresponding to relative reductions of 2.3% and 3.1%, respectively. These variants share the EPK architecture, energy coefficient, training protocol, and sampling budget. Thus, the improvement cannot be explained solely by introducing additional conditioning capacity or an energy penalty of the same form; it depends on the drug-specific physiological mapping. Furthermore, comparing the conditioning-only variant against the no-EPK baseline reduces one-step waveform MSE from 0.058438 to 0.057707 (approximately 1.25%) and one-step latent MSE from 0.050661 to 0.046782 (approximately 7.66%), supporting a clear contribution from the physiological conditioning pathway even in the absence of the energy penalty.

    Second, the energy term introduces an additional training-time inductive bias beyond physiological conditioning. With the correct pharmacological mapping retained, setting the energy coefficient to zero increases waveform MSE from 0.057402 to 0.057707. However, this difference is small, its confidence interval includes zero, and the conditioning-only variant achieves slightly lower latent MSE (0.046782 vs. 0.047120). We therefore do not claim that the energy penalty independently or consistently improves all predictive metrics.

    The energy term encourages alignment between the predicted clean ECG latent representation and a learned projection of the ODE-derived physiological template, whereas the conditioning pathway supplies the template as model input. Because the latent projection is jointly learned, the resulting distance should be interpreted as a physiology-informed regularisation objective rather than an independently validated measure of physiological constraint violations. Furthermore, we note that the original energy-weight sweep in the manuscript and this new matched ablation use different training protocols and therefore should not be interpreted as directly interchangeable estimates. Both indicate that the incremental effect of energy regularisation is modest and depends on the evaluated metric and experimental configuration, whereas our principal empirical evidence for drug-specific physiological guidance comes from the matched pharmacological, identity, and permuted mapping comparisons. Specifically, we will revise our contribution summary to state that we propose a pharmacologically structured physiological guidance mechanism for action-conditioned ECG diffusion, integrating an ODE-derived cardiac template through both model conditioning and a latent-space energy objective—where the correctness of the drug-specific physiological mapping materially improves ECG generation, while the additional energy penalty provides a smaller, metric-dependent training-time inductive bias.

    The five ablations compare the effects of the mapping, EPK conditioning, and the energy term on one-step prediction. All five variants use the same diffusion backbone, data splits, training seeds (17, 42, and 2026), 500 training epochs, and sampling budget (1,000 DDPM steps, K=1K=1), with matched random streams for sampling. The pharmacological, identity, and permuted mappings share the same EPK architecture. The conditioning-only variant sets the energy coefficient to zero. The no-EPK variant removes the EPK branch. All variants retain the drug action input.

VariantEPK conditioningλEPK\lambda_{\mathrm{EPK}}One-step latent MSE ↓One-step waveform MSE ↓
Full EPK (pharmacological mapping)Pharmacological0.10.0471200.057402
Identity mappingIdentity0.10.0473960.058772
Permuted mappingPermuted0.10.0475030.059211
EPK conditioning onlyPharmacological00.0467820.057707
No EPK moduleNone00.0506610.058438

Full EPK yielded the lowest one-step waveform MSE, and the conditioning-only variant yielded the lowest one-step latent MSE. Waveform MSE was 0.001370 lower with full EPK than with the identity mapping and 0.001809 lower than with the permuted mapping. The 95% confidence intervals for both differences exclude zero. Relative to the conditioning-only variant, full EPK reduced waveform MSE by 0.000306, with a confidence interval that includes zero.

  • Under the original evaluation protocol, full EPK yielded lower MAE than the model without EPK on all six interval endpoints. The heart rate used as a conditioning input at inference is extracted directly from the original pre-dose ECG recorded at the corresponding visit. Measured post-dose heart rate is not used as a model input. The 95% confidence intervals for five of the six differences exclude zero.
Endpoint in the original evaluationFull EPK MAENo EPK MAEFull EPK − no EPK, subject-level 95% CI
Lead II QTc27.82231.793−3.971 [−6.977, −0.754]
Lead II PR16.70717.945−1.238 [−2.107, −0.373]
Lead II TpTe10.41510.662−0.247 [−0.847, 0.258]
Lead aVF QTc27.90231.125−3.223 [−5.926, −0.639]
Lead aVF PR14.70516.541−1.836 [−2.543, −1.148]
Lead aVF TpTe10.14011.223−1.083 [−1.724, −0.407]
  • Full EPK yielded the lowest average MAE across the six interval endpoints. Interval MAE is reported in milliseconds.

    We define interval change as the interval in the target ECG minus the corresponding interval in the current ECG. Predicted and observed changes use the same current ECG as the reference, so the MAE for interval change equals the MAE for predicting the target interval. The table reports means across the three training seeds. The separate pre-dose to first post-dose subset is reported under its own pairing protocol.

EndpointFull EPKIdentity mappingPermuted mappingEPK conditioning onlyNo EPK
Lead II QTc27.37528.67327.94127.71931.654
Lead II PR16.35116.67217.51816.84218.518
Lead II TpTe10.29710.19710.04110.04110.696
Lead aVF QTc27.21229.27928.55028.19131.668
Lead aVF PR14.35514.73215.30814.75117.187
Lead aVF TpTe10.02710.20710.47610.32111.277
Mean MAE across six endpoints17.60318.29318.30617.97720.167

Full EPK yielded lower MAE than the no-EPK model across all six interval endpoints. The 95% confidence intervals for five of the six differences exclude zero.

W4 and Q3. ECG pairing, action conditioning, and scope of dose and timing

  • The model predicts a subsequent ECG from the current ECG and the recorded medication regimen. The medication input encodes a fixed regimen. Administration times are used to pair the observed ECGs.

  • In the crossover ranking evaluation, each regimen uses its own visit's ECG recorded 0.5 h before dosing. We match records across regimens by source cohort, participant, nominal post-dose time, and replicate index. Elapsed time after dosing is used only for evaluation pairing and grouping.

  • The two evaluations use different starting points. The adjacent-observation evaluation predicts the next ECG from the preceding ECG within the same visit. The baseline evaluation below takes the pre-dose ECG as input and evaluates the prediction against the first observed post-dose ECG.

  • In the pre-dose to first post-dose evaluation, full EPK achieved lower QTc-change MAE in leads II and aVF. The model without EPK had slightly lower latent MSE.

Pre-dose to first post-dose metric ↓Full EPKNo EPK
Latent MSE0.0308330.030550
Lead II QTc-change MAE (ms)23.67528.385
Lead aVF QTc-change MAE (ms)26.82430.665

W6. Bias and predictive dispersion under prior mismatch

  • Under the original one-step evaluation protocol, full EPK reduced latent MSE by 7.01% and waveform MSE by 1.78% relative to the model without EPK.
One-step metric under the original protocol ↓Full EPKNo EPK moduleRelative reduction in the mean
Latent MSE0.0471090.0506627.01%
Waveform MSE0.0574100.0584531.78%

Full EPK reduced latent MSE across all training seeds and improved QTc, PR, and TpTe errors in leads II and aVF.

  • In recursive prediction at a nominal horizon of 24 h, full EPK had the lowest endpoint latent and waveform MSEs among the five variants. The increase in error due to recursion is the extra error incurred by using the preceding predicted ECG in place of the observed ECG.
VariantNominal 24 h latent MSE ↓Nominal 24 h waveform MSE ↓
Full EPK (pharmacological mapping)0.0455280.058502
Identity mapping0.0461400.059843
Permuted mapping0.0459850.059946
EPK conditioning only0.0456310.058950
No EPK module0.0514550.059857
Reviewer_JhWmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper rightfully argues current large-scale foundation models can not really model interventions. It proposes and action conditioned diffusion model with multiple components of a VAE based latent representation, action conditioned diffusion model and also a McSharry dynamical system as a prior to tilt the distribution. It is later evaluated on a non counterfactual-way in the observational data using largely MIMIC-IV dataset.

Critical strengths
  • Problem is definetely interesting, important and relevant. Improving counterfactual treatmen analysis via generative model would be infinitely helpful to the community.

  • The prior using the physiological model is sensible and a good decision.

  • Modeling risk and uncertanity is a good addition.

Critical weaknesses
  • The model is evaluated in an observaitonal way. Main interventation claimes are not supported by a counter-factual evaluation. The eval is on a held-out subjects which are chosen randomly. It does not model/sample interventions in a counter-factual way. This is against the claim of counter-factuality in the paper. I would expect euther a randomized interventation data or some clinincally meaningful grouping and modeling.

  • The evalution is following an average of probabilities as diagnostics. Why is this relevant? Is there any clinical guidance for this choice.

  • Uncertainty is a good component but not sufficiantly validated. There is no calibration or bias analysis for the uncertainty.

  • Baselines are existing LLMs not designed for the task. They are not really meaningful comparisons. I would expect simpler baselines like simple supervised deterministic regression model or a simple transformer instead of a generative model.

Questions for authors

See strengths and weaknesses

2 comments

W1. Evaluation scope and evidence for interventions

  • We agree that prediction under an observed intervention alone does not establish counterfactual accuracy. To address this concern, we additionally evaluate intervention-response ranking using randomised crossover data. Unlike the MIMIC evaluation, these trials provide observations under multiple drug regimens for the same participants. Predictions use the corresponding visit-specific pre-dose ECG, and the evaluation compares baseline-adjusted drug responses across treatment periods. We therefore interpret this experiment as within-subject cross-intervention response validation, rather than direct identification of multiple potential outcomes from one identical baseline state.

  • We rank candidate regimens by their predicted QTc prolongation cost and compare that ranking with the ranking based on observed ECGs. The weight λ\lambda controls the contribution of the predicted cost's standard deviation. At λ=0\lambda=0, the ranking uses the predicted mean cost alone.

    ΔQTc is the change in QTc from the pre-dose baseline. For placebo-adjusted ΔΔQTc analyses, the model's predicted ranking subtracts the model-predicted ΔQTc under the matched placebo action from the candidate regimen's predicted ΔQTc, whereas the reference ranking subtracts the measured ΔQTc from the matched placebo visit from the measured candidate ΔQTc.

    Regret is the difference in observed QTc cost between the model's top-ranked regimen and the best regimen according to the observed data (ms, lower is better). Top-1 is the proportion of cases in which the model's top-ranked regimen matches the observed best regimen, accounting for ties (higher is better). Kendall τb\tau_b measures agreement between the predicted and observed rankings while accounting for ties (higher is better).

    We average each metric within participants and then give participants equal weight. We compute 95% confidence intervals with a participant-level percentile bootstrap using 10,000 resamples.

    For lead II ΔQTc without placebo adjustment, using λ=1\lambda=1 instead of the mean-only ranking reduced regret from 10.300 to 8.812 ms (paired difference −1.488 ms, 95% CI [−3.281, −0.033]) and increased Top-1 agreement from 55.62% to 58.81% (paired difference +3.19 percentage points, 95% CI [0.17, 6.16]). These results are exploratory, given the limited number of independent participants and comparisons across multiple endpoints and uncertainty weights. The observed benefit of uncertainty penalisation is endpoint-dependent and does not establish superior clinical treatment selection.

EndpointλRegret (ms) ↓Top-1 (%) ↑Kendall τb ↑
II ΔQTc010.30055.620.1426
II ΔQTc0.59.24955.400.1426
II ΔQTc18.81258.810.1815
II ΔQTc29.60357.600.1774
aVF ΔQTc010.16056.160.2642
aVF ΔQTc0.510.12055.350.2844
aVF ΔQTc110.34854.380.2726
aVF ΔQTc210.27853.900.2687
II ΔΔQTc010.62656.670.2690
II ΔΔQTc0.510.92856.730.2654
II ΔΔQTc110.83057.840.2636
II ΔΔQTc210.83057.840.2675
aVF ΔΔQTc012.13950.180.2621
aVF ΔΔQTc0.511.58650.920.2599
aVF ΔΔQTc111.70950.120.2542
aVF ΔΔQTc212.09550.120.2477
  • The clinical evaluation assesses ECG prediction under recorded treatments. The crossover evaluation compares responses to different regimens within the same participant. We use MIMIC-derived data for the clinical evaluation and ECGRDVQ/ECGDMMLD data for the crossover evaluation, which includes placebo and combination therapies. Both evaluations compare the predicted ECG with the ECG observed under the corresponding regimen.

    In the crossover evaluation, we generate each post-dose ECG in one prediction step, conditioning on the ECG recorded 0.5 h before dosing in that regimen's visit and the corresponding drug action. We match records across regimens by source cohort, participant, nominal post-dose time, and replicate index. Elapsed time after dosing is used only for evaluation pairing and grouping. For each regimen, baseline QTc is extracted directly from the original pre-dose ECG recorded at the corresponding visit. Predicted and observed QTc changes are calculated by subtracting this same pre-dose baseline from the generated and measured post-dose QTc values, respectively.

    Each ranking comparison uses a set of at least two non-placebo regimens for the same participant, nominal post-dose time, and replicate index. Placebo serves only as a reference for adjustment. The table gives the common evaluable sample sizes by endpoint.

EndpointParticipantsEvaluable ranking comparisons
Lead II ΔQTc8164
Lead aVF ΔQTc8217
Lead II placebo-adjusted ΔΔQTc592
Lead aVF placebo-adjusted ΔΔQTc5128

W2. Definition and clinical interpretation of the diagnostic proxy

  • We use the mean probability across 17 labels to compare the diagnostic outputs for generated and observed ECGs. The labels include sinus rhythm, pathological findings, and signal quality. A higher mean therefore has no uniform interpretation as clinical risk. We rank candidate regimens by positive QTc prolongation relative to the pre-dose baseline and report results with and without placebo adjustment.

W3. Predictive uncertainty, calibration, and bias

  • We agree that dispersion across generated ECGs should not be interpreted as calibrated clinical uncertainty. Our additional evaluation reveals substantial undercoverage of nominal 90% prediction intervals, with mixed CRPS results across endpoints. We therefore restrict our conclusions to empirical predictive-distribution assessment and exploratory uncertainty-aware QTc ranking, rather than calibrated clinical risk estimation.

    We compare the predictive distributions with observations using CRPS, empirical prediction-interval coverage, and interval width. The prediction intervals have 90% nominal coverage.

Predicted quantityCRPS (full EPK / no EPK)Empirical coverage (90% nominal, full EPK / no EPK)Interval width (full EPK / no EPK)
Latent dimensions0.103025 / 0.09778580.09% / 69.66%0.61641 / 0.36055
Lead II QTc25.536 / 16.245 ms63.59% / 69.74%82.812 / 68.338 ms
Lead aVF QTc16.063 / 16.475 ms75.99% / 70.63%76.923 / 66.149 ms

Full EPK increases coverage and widens the prediction intervals for the latent variables and lead aVF QTc. Latent CRPS increases by 0.005240 (95% CI [0.004611, 0.005862]), while lead aVF QTc CRPS decreases slightly. For lead II QTc, CRPS is higher and coverage is lower with full EPK. Empirical coverage falls below the nominal 90% for all three quantities.

W4. Deterministic baselines for the same prediction task

  • The Full EPK model achieves approximately 16.5% lower latent MSE than ridge regression under this reported predictive-mean comparison (0.030584 vs. 0.036645), and 37.7% lower latent MSE than the deterministic MLP (0.030584 vs. 0.049130). The predictive means of both diffusion models outperform the deterministic baselines, with full EPK achieving the lowest latent MSE. All models predict the latent representation of the next ECG in the same task. Moreover, adding the treatment regimen input to ridge regression (0.036649 vs. 0.036645) or the deterministic MLP (0.049153 vs. 0.049130) yields no improvement over state-only regression, whereas incorporating action conditioning and pharmacological guidance in Full EPK substantially reduces prediction error.
Prediction modelLatent MSE ↓
Copy the current ECG latent representation0.057359
Ridge regression (current latent representation)0.036645
Ridge regression (current latent representation + treatment regimen)0.036649
Deterministic MLP (current latent representation)0.049130
Deterministic MLP (current latent representation + treatment regimen)0.049153
Diffusion model without EPK (predictive mean)0.032755
Full EPK diffusion model (predictive mean)0.030584
  • Both ridge regression variants with shuffled treatment conditioning have similar errors, close to the error of the control using the recorded treatment conditions.
Treatment conditioning during ridge trainingTreatment conditioning during evaluationLatent MSE ↓
RecordedRecorded0.036649
ShuffledShuffled0.036710
RecordedShuffled0.036742

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.