SE-Diff: Simulator and Experience Enhanced Diffusion Model for Comprehensive ECG Generation

Xiaoda Wang; Kaiqiao Han; Yuhao Xu; Xiao Luo; Yizhou Sun; Wei Wang; Carl Yang

cs.LGNov 13, 2025

SE-Diff: Simulator and Experience Enhanced Diffusion Model for Comprehensive ECG Generation

Authors: Xiaoda Wang, Kaiqiao Han, Yuhao Xu, Xiao Luo, Yizhou Sun, Wei Wang, Carl Yang

Organizations: Emory University · University of California, Los Angeles · University of Wisconsin–Madison

Abstract

Cardiovascular disease (CVD) is a leading cause of mortality worldwide. Electrocardiograms (ECGs) are the most widely used non-invasive tool for cardiac assessment, yet large, well-annotated ECG corpora are scarce due to cost, privacy, and workflow constraints. Generating ECGs can be beneficial for the mechanistic understanding of cardiac electrical activity, enable the construction of large, heterogeneous, and unbiased datasets, and facilitate privacy-preserving data sharing. Generating realistic ECG signals from clinical context is important yet underexplored. Recent work has leveraged diffusion models for text-to-ECG generation, but two challenges remain: (i) existing methods often overlook the physiological simulator knowledge of cardiac activity; and (ii) they ignore broader, experience-based clinical knowledge grounded in real-world practice. To address these gaps, we propose SE-Diff, a novel physiological simulator and experience enhanced diffusion model for comprehensive ECG generation. SE-Diff integrates a lightweight ordinary differential equation (ODE)-based ECG simulator into the diffusion process via a beat decoder and simulator-consistent constraints, injecting mechanistic priors that promote physiologically plausible waveforms. In parallel, we design an LLM-powered experience retrieval-augmented strategy to inject clinical knowledge, providing more guidance for ECG generation. Extensive experiments on real-world ECG datasets demonstrate that SE-Diff improves both signal fidelity and text-ECG semantic alignment over baselines, proving its superiority for text-to-ECG generation. We further show that the simulator-based and experience-based knowledge also benefit downstream ECG classification.

Figures & tables

Figure 1: Overview Framework of SE-Diff. (a) Variational Autoencoder : encoder–decoder with a lightweight beat decoder for a single QRS-aligned cycle. (b) Conditional latent diffusion : U-Net denoiser with cross-attention to text, metadata, and retrieved report. (c) Simulator-informed diffusion : Euler and inter-lead constraints on the beat decoder output. (d) Experience retrieval–augmented Conditioning : tri-view EHR similarity with LLM distillation into a concise report. (e) Inference : reverse diffusion and decoding to a 10-second, 12-lead ECG.

MIMIC-IV-ECG (internal)
Model	PR $\downarrow$	QRSd $\downarrow$	QT $\downarrow$	QTcF $\downarrow$	ST@J+60 $\downarrow$	P dur $\downarrow$	T dur $\downarrow$
SSSD	$7.70\pm 8$	$32.00\pm 1$	$25.11\pm 11$	$23.25\pm 11$	$0.16\pm 0$	$5.30\pm 3$	$23.71\pm 9$
WGAN	$18.01\pm 9$	$29.64\pm 9$	$21.02\pm 5$	$23.83\pm 2$	$0.13\pm 0$	$6.01\pm 0$	$24.01\pm 4$
BeatDiff	$10.34\pm 5$	$18.23\pm 6$	$14.23\pm 6$	$9.43\pm 5$	$0.17\pm 0$	$7.00\pm 1$	$30.02\pm 3$
DiffuSETS	$14.81\pm 6$	$11.71\pm 7$	$8.20\pm 3$	$9.71\pm 4$	$0.04\pm 0$	$5.60\pm 1$	$8.71\pm 7$
SE-Diff	$\mathbf{7.30\pm 3}$	$\mathbf{10.71\pm 4}$	$\mathbf{4.50\pm 2}$	$\mathbf{7.88\pm 2}$	$\mathbf{0.03\pm 0}$	$\mathbf{2.50\pm 0}$	$\mathbf{6.80\pm 3}$

Table 2: Beat-level morphology & interval fidelity (MAE) on MIMIC-IV-ECG and PTB-XL datasets. All values are record-level medians; lower is better (

\downarrow

Model	Male = 10% Female			SVT = 10% Sinus
	F1 (%, $\uparrow$ )	Acc. (%, $\uparrow$ )	AUC (%, $\uparrow$ )	F1 (%, $\uparrow$ )	Acc. (%, $\uparrow$ )	AUC (%, $\uparrow$ )
SSSD	$42\pm 0$	$54\pm 1$	$49\pm 2$	$56\pm 1$	$63\pm 2$	$81\pm 2$
WGAN	$42\pm 0$	$54\pm 1$	$49\pm 2$	$57\pm 2$	$63\pm 2$	$82\pm 2$
BeatDiff	$44\pm 2$	$55\pm 2$	$51\pm 2$	$60\pm 2$	$67\pm 2$	$84\pm 3$
DiffuSETS	$44\pm 3$	$54\pm 2$	$54\pm 1$	$70\pm 1$	$68\pm 1$	$84\pm 2$
SE-Diff	$\mathbf{58}\pm\mathbf{1}$	$\mathbf{58}\pm\mathbf{1}$	$\mathbf{58}\pm\mathbf{2}$	$\mathbf{72}\pm\mathbf{2}$	$\mathbf{71}\pm\mathbf{0}$	$\mathbf{85}\pm\mathbf{2}$

Table 3: Downstream ECG classification under severe class imbalance.

Panel of Figure 2 — Figure 2: Noise scheduling analysis showing the progression of noise and signal factors throughout the diffusion process.

Panel of Figure 3 — Figure 3: Representative single-cycle ECG waveforms generated from our simulator. Panel A: sinus rhythm (Lead I). Panel B: ventricular pacing (Lead V1). Panel C: sinus rhythm with first-degree AV block (Lead II). Panel D: consider acute ST-elevation MI (Lead V3).

Panel of Figure 4 — Figure 4: Case Study for ECG Generation.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Prompt examples.

Model	MAE $\downarrow$	NRMSE $\downarrow$	$\mathrm{MAE}_{\text{HR}}$ $\downarrow$	rCLIP Score $\uparrow$	rFID Score $\uparrow$
SE-Diff	$\mathbf{0.0923\pm 0.0021}$	$\mathbf{0.0714\pm 0.0010}$	$\mathbf{8.43\pm 0.42}$	$\mathbf{0.9470\pm 0.0029}$	$\mathbf{0.9509\pm 0.0035}$
w/o Beat Loss	$0.0951\pm 0.0020$	$0.0743\pm 0.0016$	$23.48\pm 1.07$	$0.9393\pm 0.0026$	$0.9140\pm 0.0055$
w/o Spec Loss	$0.0942\pm 0.0023$	$0.0797\pm 0.0014$	$16.47\pm 1.23$	$0.9257\pm 0.0030$	$0.9073\pm 0.0055$

Appendix

Table 4: ECG generation performance on MIMIC-IV-ECG dataset.

Batch size	Time / batch (s)	Time / sample (s)
32	13.55	0.425
128	13.65	0.105
256	13.40	0.050
512	15.05	0.0295
1024	24.90	0.0245
2048	45.00	0.0220

Appendix

Table 5: Inference latency of the model for generating one 10-second, 12-lead ECG under different batch sizes. Times are averaged over 3 batches per configuration, with 1000 diffusion sampling steps.

Sampling Steps	NRMSE $\downarrow$	$\mathrm{MAE}_{\text{HR}}$ $\downarrow$	rCLIP Score $\uparrow$
1000	$\mathbf{0.0714\pm 0.0010}$	$\mathbf{8.43\pm 0.42}$	$\mathbf{0.9470\pm 0.0029}$
800	$0.0877\pm 0.0011$	$13.59\pm 1.05$	$0.8671\pm 0.0040$
500	$0.1097\pm 0.0010$	$15.60\pm 1.19$	$0.7888\pm 0.0084$

Appendix

Table 6: ECG generation performance under different sampling steps.

Sampler	NRMSE $\downarrow$	$\mathrm{MAE}_{\text{HR}}$ $\downarrow$	rCLIP Score $\uparrow$
SE-Diff (DDPM)	$0.0714\pm 0.0010$	$\mathbf{8.43\pm 0.42}$	$\mathbf{0.9470\pm 0.0029}$
DDIM	$0.1219\pm 0.0010$	$9.05\pm 0.44$	$0.7783\pm 0.0072$
Second Order	$0.0877\pm 0.0011$	$13.59\pm 1.05$	$0.8671\pm 0.0040$
SDE	$\mathbf{0.0693\pm 0.0106}$	$19.21\pm 1.27$	$0.9213\pm 0.0028$

Appendix

Table 7: ECG generation performance under different samplers.

Sampler	NRMSE $\downarrow$	$\mathrm{MAE}_{\text{HR}}$ $\downarrow$	rCLIP Score $\uparrow$
SE-Diff (Linear)	$\mathbf{0.0714\pm 0.0010}$	$\mathbf{8.43\pm 0.42}$	$\mathbf{0.9470\pm 0.0029}$
Cosine	$0.0728\pm 0.0032$	$8.90\pm 0.78$	$0.9446\pm 0.0131$

Appendix

Table 8: ECG generation performance under different noise schedules.

Config	Model	MAE $\downarrow$	$\mathrm{MAE}_{\text{HR}}$ $\downarrow$	rCLIP Score $\uparrow$	rFID Score $\uparrow$
A	DiffuSETS	$0.1092\pm 0.0022$	$13.29\pm 1.13$	$0.9309\pm 0.0036$	$0.9209\pm 0.0038$
A	SE-Diff	$\mathbf{0.0923\pm 0.0021}$	$\mathbf{8.43\pm 0.42}$	$\mathbf{0.9470\pm 0.0029}$	$\mathbf{0.9509\pm 0.0035}$
B	DiffuSETS	$0.1092\pm 0.0022$	$13.29\pm 1.13$	$0.9347\pm 0.0032$	$0.9197\pm 0.0037$
B	SE-Diff	$\mathbf{0.0923\pm 0.0021}$	$\mathbf{8.43\pm 0.42}$	$\mathbf{0.9472\pm 0.0032}$	$\mathbf{0.9505\pm 0.0035}$
C	DiffuSETS	$0.1092\pm 0.0022$	$13.29\pm 1.13$	$0.9242\pm 0.0039$	$0.9209\pm 0.0038$
C	SE-Diff	$\mathbf{0.0923\pm 0.0021}$	$\mathbf{8.43\pm 0.42}$	$\mathbf{0.9393\pm 0.0031}$	$\mathbf{0.9509\pm 0.0035}$

Appendix

Table 9: ECG generation performance comparison between SE-Diff and DiffuSETS under different calibration configurations.

$\boldsymbol{\lambda_{1}}$	$\boldsymbol{\lambda_{2}}$	$\boldsymbol{\lambda_{3}}$	$\boldsymbol{k}$	LLM Score
1	1	1	3	$\mathbf{94.9\pm 2.3}$
1	1	1	2	$94.4\pm 2.1$
1	1	1	4	$94.8\pm 2.9$
2	1	1	3	$94.7\pm 2.6$
1	2	1	3	$93.6\pm 1.5$
1	1	2	3	$93.5\pm 3.3$

Appendix

Table 10: Sensitivity of tri-view weights

(\lambda_{1},\lambda_{2},\lambda_{3})

and top-

k

on retrieval performance.

Synthetic Clinical Data Generation

See all 27 papers

ECG-WM: A Physiology-Informed ECG World Model for Clinical Intervention Simulation
May 17, 2026Zhikang Chen, Yue Wang, Sen Cui +4Conditional Diffusion Models Clinical Risk Prediction
University of Oxford · Tsinghua University · Southern University of Science and Technology
CAMEL: An ECG Language Model for Forecasting Cardiac Events
Feb 17, 2026Neelay Velingker, Alaia Solko-Breslin, Mayank Keoliya +9Electrocardiogram Foundation Models Electrocardiogram Interpretation
University of Pennsylvania
ECGFlowCMR: Pretraining with ECG-Generated Cine CMR Helps Cardiac Disease Classification and Phenotype Prediction
Jan 28, 2026Xiaocheng Fang, Zhengyao Ding, Guangkun Nie +9Synthetic Data Pretraining Medical Diagnosis
State Key Laboratory of General Artificial Intelligence, Peking University · School of Intelligence Science and Technology, Peking University, Beijing, China · Polytechnic Institute of Zhejiang University, Zhejiang University, Hangzhou, China +4

May 17, 2026J/K move · Enter open · S save

ECG-WM: A Physiology-Informed ECG World Model for Clinical Intervention Simulation

Zhikang Chen, Yue Wang, Sen Cui, Yu Zhang+3

University of Oxford · Tsinghua University · Southern University of Science and Technology

Conditional Diffusion Models Clinical Risk Prediction

Open paper AskPDF arXiv

Electrocardiogram (ECG)-based models have achieved strong performance in diagnostic tasks, yet they remain limited in modeling how cardiac dynamics evolve under external interventions. In particular, existing approaches focus primarily on static prediction and lack mechanisms to capture ECG variations under different pharmacological conditions. In this work, we propose an ECG World Model for action-conditioned predictive simulation of cardiac electrophysiology. Moving beyond disjoint pipelines, our framework features a principled integration of physiological ordinary differential equation (ODE) priors into latent diffusion dynamics via energy regularization. This structural constraint enables the synthesis of physiologically plausible post-intervention ECG trajectories while effectively mitigating generative hallucinations. Building on this simulation process, we introduce an uncertainty-aware evaluation strategy that leverages the stochasticity of diffusion sampling to characterize both the expected clinical risk and its variability, allowing a more reliable comparative assessment of candidate interventions. We evaluate our method across diverse settings, including controlled drug-response scenarios and real-world clinical records. Beyond standard waveform metrics, experimental results demonstrate improved risk calibration and strong alignment with expert-informed treatment preferences. These results establish our approach as a robust foundation for safe and intervention-aware clinical decision support.

Abstract

Synthetic Clinical Data Generation

Explore similar work

ECG-WM: A Physiology-Informed ECG World Model for Clinical Intervention Simulation

CAMEL: An ECG Language Model for Forecasting Cardiac Events

ECGFlowCMR: Pretraining with ECG-Generated Cine CMR Helps Cardiac Disease Classification and Phenotype Prediction