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

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

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.
ModelPR ↓\downarrowQRSd ↓\downarrowQT ↓\downarrowQTcF ↓\downarrowST@J+60 ↓\downarrowP dur ↓\downarrowT dur ↓\downarrow
MIMIC-IV-ECG (internal)
SSSD7.70±87.70\pm 832.00±132.00\pm 125.11±1125.11\pm 1123.25±1123.25\pm 110.16±00.16\pm 05.30±35.30\pm 323.71±923.71\pm 9
WGAN18.01±918.01\pm 929.64±929.64\pm 921.02±521.02\pm 523.83±223.83\pm 20.13±00.13\pm 06.01±06.01\pm 024.01±424.01\pm 4
BeatDiff10.34±510.34\pm 518.23±618.23\pm 614.23±614.23\pm 69.43±59.43\pm 50.17±00.17\pm 07.00±17.00\pm 130.02±330.02\pm 3
DiffuSETS14.81±614.81\pm 611.71±711.71\pm 78.20±38.20\pm 39.71±49.71\pm 40.04±00.04\pm 05.60±15.60\pm 18.71±78.71\pm 7
SE-Diff7.30±3\mathbf{7.30\pm 3}10.71±4\mathbf{10.71\pm 4}4.50±2\mathbf{4.50\pm 2}7.88±2\mathbf{7.88\pm 2}0.03±0\mathbf{0.03\pm 0}2.50±0\mathbf{2.50\pm 0}6.80±3\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 ).
ModelMale = 10% FemaleSVT = 10% Sinus
 F1 (%, ↑\uparrow )Acc. (%, ↑\uparrow )AUC (%, ↑\uparrow )F1 (%, ↑\uparrow )Acc. (%, ↑\uparrow )AUC (%, ↑\uparrow )
SSSD42±042\pm 054±154\pm 149±249\pm 256±156\pm 163±263\pm 281±281\pm 2
WGAN42±042\pm 054±154\pm 149±249\pm 257±257\pm 263±263\pm 282±282\pm 2
BeatDiff44±244\pm 255±255\pm 251±251\pm 260±260\pm 267±267\pm 284±384\pm 3
DiffuSETS44±344\pm 354±254\pm 254±154\pm 170±170\pm 168±168\pm 184±284\pm 2
SE-Diff58±1\mathbf{58}\pm\mathbf{1}58±1\mathbf{58}\pm\mathbf{1}58±2\mathbf{58}\pm\mathbf{2}72±2\mathbf{72}\pm\mathbf{2}71±0\mathbf{71}\pm\mathbf{0}85±2\mathbf{85}\pm\mathbf{2}
Table 3: Downstream ECG classification under severe class imbalance.
Figure 2: Noise scheduling analysis showing the progression of noise and signal factors throughout the diffusion process.
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).
Figure 4: Case Study for ECG Generation.
Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Prompt examples.
ModelMAE ↓\downarrowNRMSE ↓\downarrowMAEHR\mathrm{MAE}_{\text{HR}} ↓\downarrowrCLIP Score ↑\uparrowrFID Score ↑\uparrow
SE-Diff0.0923±0.0021\mathbf{0.0923\pm 0.0021}0.0714±0.0010\mathbf{0.0714\pm 0.0010}8.43±0.42\mathbf{8.43\pm 0.42}0.9470±0.0029\mathbf{0.9470\pm 0.0029}0.9509±0.0035\mathbf{0.9509\pm 0.0035}
w/o Beat Loss0.0951±0.00200.0951\pm 0.00200.0743±0.00160.0743\pm 0.001623.48±1.0723.48\pm 1.070.9393±0.00260.9393\pm 0.00260.9140±0.00550.9140\pm 0.0055
w/o Spec Loss0.0942±0.00230.0942\pm 0.00230.0797±0.00140.0797\pm 0.001416.47±1.2316.47\pm 1.230.9257±0.00300.9257\pm 0.00300.9073±0.00550.9073\pm 0.0055
Appendix
Table 4: ECG generation performance on MIMIC-IV-ECG dataset.
Batch sizeTime / batch (s)Time / sample (s)
3213.550.425
12813.650.105
25613.400.050
51215.050.0295
102424.900.0245
204845.000.0220
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 StepsNRMSE ↓\downarrowMAEHR\mathrm{MAE}_{\text{HR}} ↓\downarrowrCLIP Score ↑\uparrow
10000.0714±0.0010\mathbf{0.0714\pm 0.0010}8.43±0.42\mathbf{8.43\pm 0.42}0.9470±0.0029\mathbf{0.9470\pm 0.0029}
8000.0877±0.00110.0877\pm 0.001113.59±1.0513.59\pm 1.050.8671±0.00400.8671\pm 0.0040
5000.1097±0.00100.1097\pm 0.001015.60±1.1915.60\pm 1.190.7888±0.00840.7888\pm 0.0084
Appendix
Table 6: ECG generation performance under different sampling steps.
SamplerNRMSE ↓\downarrowMAEHR\mathrm{MAE}_{\text{HR}} ↓\downarrowrCLIP Score ↑\uparrow
SE-Diff (DDPM)0.0714±0.00100.0714\pm 0.00108.43±0.42\mathbf{8.43\pm 0.42}0.9470±0.0029\mathbf{0.9470\pm 0.0029}
DDIM0.1219±0.00100.1219\pm 0.00109.05±0.449.05\pm 0.440.7783±0.00720.7783\pm 0.0072
Second Order0.0877±0.00110.0877\pm 0.001113.59±1.0513.59\pm 1.050.8671±0.00400.8671\pm 0.0040
SDE0.0693±0.0106\mathbf{0.0693\pm 0.0106}19.21±1.2719.21\pm 1.270.9213±0.00280.9213\pm 0.0028
Appendix
Table 7: ECG generation performance under different samplers.
SamplerNRMSE ↓\downarrowMAEHR\mathrm{MAE}_{\text{HR}} ↓\downarrowrCLIP Score ↑\uparrow
SE-Diff (Linear)0.0714±0.0010\mathbf{0.0714\pm 0.0010}8.43±0.42\mathbf{8.43\pm 0.42}0.9470±0.0029\mathbf{0.9470\pm 0.0029}
Cosine0.0728±0.00320.0728\pm 0.00328.90±0.788.90\pm 0.780.9446±0.01310.9446\pm 0.0131
Appendix
Table 8: ECG generation performance under different noise schedules.
ConfigModelMAE ↓\downarrowMAEHR\mathrm{MAE}_{\text{HR}} ↓\downarrowrCLIP Score ↑\uparrowrFID Score ↑\uparrow
ADiffuSETS0.1092±0.00220.1092\pm 0.002213.29±1.1313.29\pm 1.130.9309±0.00360.9309\pm 0.00360.9209±0.00380.9209\pm 0.0038
ASE-Diff0.0923±0.0021\mathbf{0.0923\pm 0.0021}8.43±0.42\mathbf{8.43\pm 0.42}0.9470±0.0029\mathbf{0.9470\pm 0.0029}0.9509±0.0035\mathbf{0.9509\pm 0.0035}
BDiffuSETS0.1092±0.00220.1092\pm 0.002213.29±1.1313.29\pm 1.130.9347±0.00320.9347\pm 0.00320.9197±0.00370.9197\pm 0.0037
BSE-Diff0.0923±0.0021\mathbf{0.0923\pm 0.0021}8.43±0.42\mathbf{8.43\pm 0.42}0.9472±0.0032\mathbf{0.9472\pm 0.0032}0.9505±0.0035\mathbf{0.9505\pm 0.0035}
CDiffuSETS0.1092±0.00220.1092\pm 0.002213.29±1.1313.29\pm 1.130.9242±0.00390.9242\pm 0.00390.9209±0.00380.9209\pm 0.0038
CSE-Diff0.0923±0.0021\mathbf{0.0923\pm 0.0021}8.43±0.42\mathbf{8.43\pm 0.42}0.9393±0.0031\mathbf{0.9393\pm 0.0031}0.9509±0.0035\mathbf{0.9509\pm 0.0035}
Appendix
Table 9: ECG generation performance comparison between SE-Diff and DiffuSETS under different calibration configurations.
λ1\boldsymbol{\lambda_{1}}λ2\boldsymbol{\lambda_{2}}λ3\boldsymbol{\lambda_{3}}k\boldsymbol{k}LLM Score
111394.9±2.3\mathbf{94.9\pm 2.3}
111294.4±2.194.4\pm 2.1
111494.8±2.994.8\pm 2.9
211394.7±2.694.7\pm 2.6
121393.6±1.593.6\pm 1.5
112393.5±3.393.5\pm 3.3
Appendix
Table 10: Sensitivity of tri-view weights (λ1,λ2,λ3)(\lambda_{1},\lambda_{2},\lambda_{3}) and top- kk on retrieval performance.
More in this topic

Synthetic Clinical Data Generation

See all 27 papers

Explore similar work

CardsList
May 17, 2026

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

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.

Zhikang Chen, Yue Wang, Sen Cui +4

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

Conditional Diffusion ModelsClinical Risk Prediction
AskPDF
Feb 17, 2026

CAMEL: An ECG Language Model for Forecasting Cardiac Events

Electrocardiograms (ECG) are electrical recordings of the heart that are critical for diagnosing cardiovascular conditions. ECG language models (ELMs) have recently emerged as a promising framework for ECG classification accompanied by report generation. However, current models cannot forecast future cardiac events despite the immense clinical value for planning earlier intervention. To address this gap, we propose CAMEL, the first ELM that is capable of inference over longer signal durations which enables its forecasting capability. Our key insight is a specialized ECG encoder which enables cross-understanding of ECG signals with text. We train CAMEL using established LLM training procedures, combining LoRA adaptation with a curriculum learning pipeline. Our curriculum includes ECG classification, metrics calculations, and multi-turn conversations to elicit reasoning. CAMEL demonstrates strong zero-shot performance across 6 tasks and 9 datasets, including ECGForecastBench, a new benchmark that we introduce for forecasting arrhythmias. CAMEL is on par with or surpasses ELMs and fully supervised baselines both in- and out-of-distribution, achieving SOTA results on ECGBench (+7.0% absolute average gain) as well as ECGForecastBench (+12.4% over fully supervised models and +21.1% over zero-shot ELMs).

Neelay Velingker, Alaia Solko-Breslin, Mayank Keoliya +9

University of Pennsylvania

Electrocardiogram Foundation ModelsElectrocardiogram InterpretationClinical Outcome PredictionElectrocardiogram ClassificationForecasting Benchmarks
AskPDF
Jan 28, 2026

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

Cardiac Magnetic Resonance (CMR) imaging provides a comprehensive assessment of cardiac structure and function but remains constrained by high acquisition costs and reliance on expert annotations, limiting the availability of large-scale labeled datasets. In contrast, electrocardiograms (ECGs) are inexpensive, widely accessible, and offer a promising modality for conditioning the generative synthesis of cine CMR. To this end, we propose ECGFlowCMR, a novel ECG-to-CMR generative framework that integrates a Phase-Aware Masked Autoencoder (PA-MAE) and an Anatomy-Motion Disentangled Flow (AMDF) to address two fundamental challenges: (1) the cross-modal temporal mismatch between multi-beat ECG recordings and single-cycle CMR sequences, and (2) the anatomical observability gap due to the limited structural information inherent in ECGs. Extensive experiments on the UK Biobank and a proprietary clinical dataset demonstrate that ECGFlowCMR can generate realistic cine CMR sequences from ECG inputs, enabling scalable pretraining and improving performance on downstream cardiac disease classification and phenotype prediction tasks.

Xiaocheng Fang, Zhengyao Ding, Guangkun Nie +9

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

Synthetic Data PretrainingMedical DiagnosisClinical PredictionMedical Image SynthesisConditional Generative ModelingCardiac MRI
AskPDF