Research map / Image generation / Fast diffusion sampling
Region-Disentangled Diffusion Model for High-Fidelity PPG-to-ECG Translation
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.
Proposes RDDM, a diffusion model for PPG-to-ECG translation that adds noise selectively to QRS regions and disentangles denoising of ROI and non-ROI parts, and introduces CardioBench, a 5-task downstream benchmark for generated ECG.
Problem
Generating high-fidelity ECG from wearable PPG signals so that the generated ECG is useful for cardiac monitoring, since standard diffusion models add noise uniformly and miss QRS morphology and temporal dynamics.
Approach
A ROI-guided forward process that adds noise only to QRS windows (defined by R-peak masks), paired with a reverse process that uses two networks: epsilon_theta denoises the ROI and rho_phi denoises the non-ROI regions, enabling high-quality generation in 10 sampling steps.
Key results
- On BIDMC, RDDM reduces RMSE from 0.63 (prior SOTA) to 0.24 and outperforms DDPM with 50 steps (0.29) using only 10 steps.
- On AFib detection (MIMIC-AFib), RDDM achieves F1 0.71 and Acc 0.65, versus 0.60 F1 and 0.43 Acc for DDPM (T=50).
- On diabetes detection (PPG-BP), RDDM reaches Acc 0.80 and F1 0.52, compared with 0.65 Acc and 0.44 F1 for the best PPG-based prior method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RDDM instead of DDPM (T = 50) | PPG-to-ECG translation · BIDMC | RMSE: 0.29 -> 0.24 | Helped |
| Use RDDM instead of DDPM (T = 50) | PPG-to-ECG translation · WESAD | FD: 9.17 -> 3.93 | Helped |
| Use RDDM instead of DDPM (T = 50) | HR estimation from generated ECG · WESAD | MAE: 4.67 -> 1.40 | Helped |
| Use RDDM instead of DDPM (T = 50) | AFib detection · MIMIC-AFib | F1: 0.60 -> 0.71 | Helped |
| Use RDDM instead of DDPM (T = 50) | Stress and affect classification · WESAD | F1: 0.58 -> 0.64 | Helped |
| Use RDDM instead of DDPM (T=50) | Diabetes detection · PPG-BP | Accuracy: 0.47 -> 0.80 | Helped |
| Use RDDM-generated ECG with PPG instead of DDPM-generated ECG with PPG | Blood pressure estimation · Cuffless-BP | MAE-SBP: 5.95 -> 2.48 | Helped |
| Use RDDM-generated ECG with PPG instead of DDPM-generated ECG with PPG | Blood pressure estimation · Cuffless-BP | MAE-DBP: 0.63 -> 1.20 | Hurt |
| Increase ROI window size from 16 to 32 (ablation) | RDDM · AFib detection | performance | Helped |
| Increase ROI window size from 32 to 64 (ablation) | RDDM · AFib detection | performance | Hurt |
| Increase sampling steps from 5 to 10 (ablation) | RDDM · HR estimation on WESAD and DALIA | performance | Helped |
| Increase sampling steps beyond 10 (ablation) | RDDM · HR estimation on WESAD and DALIA | performance | Hurt |
Surprising findings
- RDDM's SBP MAE (2.48) is lower than the upper bound using original PPG plus original ECG (3.31), which the authors attribute to noisy original ECG.
- The DDPM variant achieves the lowest DBP MAE (0.63) despite poor SBP estimation.
Limitations
- Gains on some tasks are modest or sensitive to ROI window and sampling-step hyperparameters.
- Diabetes detection trains only on generated ECG and uses a small test set, limiting statistical strength.
Experimental setup
Datasets: WESAD DALIA CAPNO BIDMC MIMIC-AFib PPG-BP Cuffless BP
Benchmarks: CardioBench WESAD CAPNO DALIA BIDMC MIMIC
Metrics: RMSE Fréchet Distance MAE F1 Accuracy
Baselines: CardioGAN DDPM (T=10/25/50) Performer Feature selection / linear regression / dictionary learning PPG-ECG methods
Builds on: DDPM CardioGAN CycleGAN
Closest papers by meaning
- PhysDiff: Physiology-based Dynamicity Disentangled Diffusion Model for Remote Physiological Measurement (AAAI 2025 · similarity 0.88)
- Leveraging an ECG Beat Diffusion Model for Morphological Reconstruction from Indirect Signals (NeurIPS 2024 · similarity 0.86)
- Constraint Latent Space Matters: An Anti-anomalous Waveform Transformation Solution from Photoplethysmography to Arterial Blood Pressure (AAAI 2024 · similarity 0.84)
- PaPaGei: Open Foundation Models for Optical Physiological Signals (ICLR 2025 · similarity 0.84)
- H-Tuning: Toward Low-Cost and Efficient ECG-based Cardiovascular Disease Detection with Pre-Trained Models (ICML 2025 · similarity 0.84)
- Cylindrical Geodesic Flow Matching for Quasiperiodic Physiological Signal Transformation (NeurIPS 2026 · 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.