Research map / Image generation / Image restoration
Leveraging an ECG Beat Diffusion Model for Morphological Reconstruction from Indirect Signals
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces BeatDiff, a lightweight 12-lead heartbeat diffusion prior, and EM-BeatDiff, an MCGDiff-based Monte Carlo EM method that solves ECG inverse problems (artifact removal, missing-lead reconstruction, QT prediction, anomaly detection) without task-specific training.
Problem
Heartbeat morphology analysis is hampered by noise and artifacts, missing leads, and scarce annotated data; the paper targets reconstructing clean 12-lead beat morphology from indirect or partial observations.
Approach
Train a small conditional denoising diffusion model on single-beat 12-lead ECGs and use it as a prior in a Bayesian linear inverse problem, sampling posteriors with sequential Monte Carlo guidance (MCGDiff) and estimating unknown observation parameters jointly via Monte Carlo EM, with no fine-tuning per task.
Key results
- BeatDiff (42 MB, ~160 ms inference) matches or beats SSSD (39 GB-scale model, 75 s inference) on sex-balancing classifier augmentation, e.g., F1 78% vs 76% at F=10%M.
- EM-BeatDiff reduces SSD for baseline wander removal to 0.14 vs 4.37 for DeScoD and achieves AUC 84.82 (MI), 93.06 (LAD), 84.73 (LQT) vs AAE 80.23, 82.69, 70.96.
- Missing-lead reconstruction of V1-6 from limb leads and smartwatch (lead I) task reaches lower SSD than EkGAN (1.03 vs 1.63 on smartwatch) while requiring no task-specific training.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BeatDiff-generated beats to balance the training data | sex classifier · sex-imbalanced classification, F = 10%M | F1: 76 ± 0 -> 78 ± 0 | Helped |
| use EM-BeatDiff instead of DeScoD for artifact removal | ECG artifact removal · Baseline Wander | SSD: 4.37 ± 8.19 -> 0.14 ± 0.01 | Helped |
| use EM-BeatDiff instead of DeScoD for artifact removal | ECG artifact removal · Electrode Motion | Cos.: 92.73 ± 0.25 -> 95.42 ± 0.19 | Helped |
| use EM-BeatDiff instead of EkGAN for missing-lead reconstruction | ECG missing-lead reconstruction · Smartwatch | SSD: 1.63 ± 0.47 -> 1.03 ± 0.05 | Helped |
| use EM-BeatDiff instead of EkGAN for missing-lead reconstruction | ECG missing-lead reconstruction · Smartwatch | Cos.: 91.33 ± 0.38 -> 86.02 ± 0.99 | Hurt |
| use EM-BeatDiff instead of EkGAN for missing-lead reconstruction | ECG missing-lead reconstruction · V1–6 reconstruction | SSD: 2.10 ± 0.83 -> 1.10 ± 0.06 | Helped |
| use EM-BeatDiff instead of AAE for anomaly detection | ECG anomaly detection · LAD | AUC: 82.69 -> 93.06 ± 0.03 | Helped |
| condition on patient age, sex, and RR (ablation) | BeatDiff · ECG generation | EMD | Helped |
| use a deeper denoiser network (ablation) | BeatDiff · ECG generation | EMD | No clear effect |
| condition MI anomaly detection on QRS rather than limb leads (ablation) | EM-BeatDiff · MI anomaly detection | AUC: 0.842 ± 0.001 -> 0.854 ± 0.001 | Helped |
| condition LAD anomaly detection on QRS rather than limb leads (ablation) | EM-BeatDiff · LAD anomaly detection | AUC: 0.908 ± 0.001 -> 0.927 ± 0.001 | Helped |
| condition LAE anomaly detection on ST rather than limb leads (ablation) | EM-BeatDiff · LAE anomaly detection | AUC: 0.759 ± 0.002 -> 0.790 ± 0.002 | Helped |
Surprising findings
- Generated T-waves under EM-BeatDiff follow the Fridericia QT-RR formula (R2 0.98) without explicitly encoding that relationship during training.
Limitations
- Trained only on healthy ECGs; arrhythmic and pathological generalization is limited.
- Multi-beat 10-second ECG reconstruction is harder and needs larger models and more particles.
- Risk of hallucination; clinical validation not performed.
Experimental setup
Datasets: PhysioNet Challenge ECG dataset MIT-BIH Noise Stress Test database MIT-BIH arrhythmia-derived noise
Benchmarks: PhysioNet Challenge ECG (sex classification) MIT-BIH Noise Stress Test (artifact removal) missing lead reconstruction (smartwatch, V1-6) anomaly detection on MI, LAD, LAE, LQT conditions
Metrics: F1 accuracy AUC SSD MAD cosine similarity EMD R2 for QT OOD score
Baselines: SSSD WGAN DeScoD EkGAN AAE
Builds on: MCGDiff EDM (Karras et al. diffusion preconditioning) Score-based ECG denoising (DeScoD) SSSD ECG generation
Closest papers by meaning
- Region-Disentangled Diffusion Model for High-Fidelity PPG-to-ECG Translation (AAAI 2024 · similarity 0.86)
- Variational Diffusion Posterior Sampling with Midpoint Guidance (ICLR 2025 · similarity 0.84)
- ECHOPulse: ECG Controlled Echocardio-gram Video Generation (ICLR 2025 · similarity 0.84)
- Parallel-Learning of Invariant and Tempo-variant Attributes of Single-Lead Cardiac Signals: PLITA (AAAI 2025 · similarity 0.84)
- Reading Your Heart: Learning ECG Words and Sentences via Pre-training ECG Language Model (ICLR 2025 · similarity 0.84)
- Boosting Masked ECG-Text Auto-Encoders as Discriminative Learners (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.