PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation
summary
The gist
Electrocardiography (ECG) is the clinical standard for cardiac assessment but requires dedicated hardware that does not scale to daily-life monitoring PPG is ubiquitous in wearables but lacks…
In short
This work addresses the gap between wearable PPG signals and clinical ECG standards by developing a physiological-guided latent rectified flow framework. It introduces an electro-hemodynamic simulator to model both ECG and PPG dynamics, using this knowledge to train a representation learning model. The resulting system generates highly accurate, physiologically plausible ECG waveforms from noisy PPG data.
Key concepts
- Electro-hemodynamic Simulator
- This is a tool that mimics the physical relationship between blood flow (hemodynamics) and electrical signals (ECG). It models how changes in heart function affect both the pulse wave seen in PPG and the resulting electrical signal recorded by an ECG, allowing the model to learn consistent dynamics across both modalities.
- Physiology-Aware AutoEncoder
- This component learns a structured latent space by using simulator guidance. It takes raw PPG and ECG data as input and maps them into a shared latent space that respects known physiological constraints, such as expected timing delays between the two signals, ensuring the learned representations are physically meaningful.
- Latent Rectified Flow
- This is a generative modeling technique used to transform random noise into structured data. In this paper, it is conditioned on PPG input and guided by simulator residuals. This ensures that the generated ECGs not only look like real signals but also follow the correct physiological timing and morphology dictated by the learned latent space.
Terminology used across episodes
This episode discusses
- PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation · Paper Radio
- PPGFlowECG: Latent Rectified Flow with Cross-Modal Encoding for PPG-Guided ECG Generation and Cardiovascular Disease Detection · Paper Radio
- Playing Atari with Deep Reinforcement Learning
- Universal Differential Equations for Scientific Machine Learning
The paper
PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation · Read on arXiv
Xiaoda Wang, Minxiao Wang, Kaiqiao Han, Defu Cao, Ching Chang, Yidan Shi, Runze Yan, Xiao Luo, Yan Liu, Xiao Hu
Department of Computer Science, Emory University Department of Computer Science, University of California, Los Angeles Department of Computer Science, University of Southern California Department of Statistics, University of Wisconsin–Madison
Electrocardiography (ECG) is the clinical standard for cardiac assessment but requires dedicated hardware that does not scale to daily-life monitoring. Photoplethysmography (PPG) is ubiquitous in wearables but lacks ECG-specific diagnostic morphology and is corrupted by motion and sensor noise. PPG-to-ECG generation aims to bridge this gap by recovering electrical morphology and timing from peripheral pulse signals. However, existing methods largely rely on statistical alignment and data-driven generation. They fail to explicitly structure the latent space around physiology-aware electro-hemodynamic factors and lack constraints from forward physiological dynamics. To address these challenges, we propose PPG2ECG, a physiology-guided latent rectified flow framework. PPG2ECG introduces an electro-hemodynamic simulator that co-models ECG and PPG through shared cardiac phase dynamics. Guided by this simulator, a Physiology-Aware AutoEncoder learns a structured electro-hemodynamic latent space. Then we integrate this simulator guidance into a PPG-conditioned latent rectified flow, enforcing ECG-side morphology consistency and ECG-to-PPG forward hemodynamic consistency during generative transport. Experiments on the large-scale MC-MED dataset demonstrate that PPG2ECG significantly improves PPG-to-ECG generation and downstream cardiovascular disease classification, proving its ability to generate ECGs that are both signal-faithful and physiologically plausible under the ECG-to-PPG hemodynamic pathway
Transcript
Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.
Rosa: We're looking at the paper: "PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation".
Dev: Here's the gist of the paper. Electrocardiography (ECG) is the clinical standard for cardiac assessment but requires dedicated hardware that does not scale to daily-life monitoring PPG is ubiquitous in wearables but lacks ECG-specific diagnostic morphology and is corrupted by motion and sensor noise PPG-to-ECG generation aims to bridge this gap by recovering electrical morphology and timing from peripheral pulse signals.
Rosa: Pruning redundant reasoning tokens PG-LRF, a physiology-guided latent rectified flow framework, introduces an electro-hemodynamic simulator that co-models ECG and PPG through shared cardiac phase dynamics.
Dev: This framework is guided by this simulator to learn a structured electro-hemodynamic latent space using a Physiology-Aware AutoEncoder.
Rosa: Subsequently, it integrates this simulator guidance into a PPG-conditioned latent rectified flow, enforcing ECG-side morphology consistency and ECG-to-PPG forward hemodynamic consistency during generative transport.
Paper summary: Rosa: So, this paper is called "PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation". It basically tackles the problem of using simple wearable pulse data, or PPG, to generate an electrocardiogram because ECG is the gold standard for clinical cardiac assessment but it needs dedicated hardware that doesn't scale well for daily use <ref:2605.12541#pg1>.
Dev: Right. The core idea here is to bridge that gap by recovering the electrical shape and timing from just a peripheral pulse signal, PPG, and they argue that current methods mostly rely on just statistical alignment and data-driven generation which doesn't explicitly structure the latent space around how the physiology actually works <ref:2605.12541#pg1>.
Rosa: Exactly. They point out two main challenges: first, there's a lack of physiology-aware latent abstraction, meaning PPG and ECG are coupled but they show different things; ECG shows electrical activation, while PPG captures the delayed pulse <ref:2605.12541#pg2>.
Dev: And the second challenge is missing simulator guidance. They say existing methods often hallucinate physically impossible delays by ignoring the actual cardiac cycle dynamics and that forward flow from ECG to PPG thirty-one twenty-four <ref:2605.12541#pg2>.
Rosa: So they propose PG-LRF as a framework for this, using an electro-hemodynamic simulator that co-models both ECG and PPG through shared cardiac phase dynamics <ref:2605.12541#pg3>. They use this simulator to guide a Physiology Aware AutoEncoder, or PAE, which learns a structured latent space by enforcing consistency between the two modalities and using latency consistency derived from the simulator <ref:2605.12541#pg3>.
Dev: And then they use this guidance in a PPG conditioned latent rectified flow to generate ECG latents from noise, where the simulator residuals steer that transport toward matching ECG morphology and ensuring forward hemodynamic consistency between ECG and PPG during training <ref:2605.12541#pg3>. It’s using the simulator as soft, phase-aware guidance instead of a hard constraint which they think avoids severe model mismatch <ref:2605.12541#pg2>.
Rosa: That means they are trying to keep the flexibility of data-driven latent generation while anchoring those generated distributions to dynamics that are electro-hemodynamically plausible <ref:2605.12541#pg3>. It’s a way to get signal fidelity while respecting the underlying physics of how the heart works <ref:2605.12541#pg3>.
Dev: From an engineering standpoint, I'm interested in how they handle that coupling between upstream electrical activation and the delayed peripheral pulse <ref:2605.12541#pg2>. How do you make sure the latent space respects that directional relationship?
Paper summary: Rosa: Well, they use a shared cardiac phase oscillator within their simulator to co-model those dynamics <ref:2605.12541#pg3>. This helps regularize the latent space using the pulse arrival delay, or delta PAT, which acts as a regularization term for the PAE latent space <ref:2605.12541#pg3>.
Dev: That makes sense for structuring the manifold. But when you're training that flow, how do you actually enforce that ECG-to-PPG forward hemodynamic consistency during the transport process? Is it just adding a loss term?
Rosa: They use two things there. First, simulator derived residuals are used to penalize deviations from the desired ECG vector field by the Euler residual <ref:2605.12541#pg3>. Second, they use PPG simulator losses to regularize generated ECGs through their induced PPG response <ref:2605.12541#pg3>.
Dev: So those simulator constraints act as soft physiological priors, narrowing down the possibilities in that underdetermined problem of going from PPG back to ECG <ref:2605.12541#pg3>. That sounds like a clever way to manage the uncertainty without over-constraining the whole thing <ref:2605.12541#pg3>.
Rosa: And they combine that statistical transport with physiological regularization to produce ECGs that are signal-faithful and preserve critical clinical intervals <ref:2605.12541#pg3>. It’s aiming for better cardiovascular disease classification on the large scale MC-MED dataset <ref:2605.12541#pg3>.
Dev: So, looking at the overall paper, "PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation", it really proposes integrating mechanistic physiological priors with data-driven latent generation <ref:2605.12541#pg2>.
Rosa: It’s about using that simulator knowledge as soft, phase aware guidance while keeping the flexibility of the data driven latent generation <ref:2605.12541#pg3>. The authors are showing how this works by introducing three components: the simulator, the PAE learning a shared latent space, and then the rectified flow guided by those residuals <ref:2605.12541#pg3>.
Dev: I think what matters is that they show improvements over previous work when you look at their metrics. They reduced Mean Absolute Error from zero point eight seven to zero point seven one and Root Mean Square Error from one point three two to one point zero seven <ref:2605.12541#pg3>. That’s a measurable improvement in waveform fidelity and accuracy compared to strong baselines <ref:2605.12541#pg3>.
Rosa: And they also saw gains in clinical error measures, like QRS error dropping from forty-eight point one to thirty-eight point six milliseconds and QTcF error from thirty-five point nine to thirty-two point two milliseconds <ref:2605.12541#pg3>. That means the generated ECGs are not just looking visually right, but they are capturing the important clinical timing intervals better than before <ref:2605.12541#pg3>.
Paper summary: Dev: The paper concludes that PPG-to-ECG generation is improved by this framework because it handles complex cross modal variations better <ref:2605.12541#pg3>. It moves beyond just data driven similarity to explicitly encode the directional physiological coupling <ref:2605.12541#pg2>.
Rosa: Exactly. So, the main implication is that this framework can help bridge the gap between ubiquitous wearable sensing and verifiable clinical cardiac assessment <ref:2605.12541#pg3>. It shows that integrating simulator knowledge as soft guidance can make data driven models much more robust for tasks like long term monitoring or when ECG isn't available <ref:2605.12541#pg3>.
Dev: It does have limitations, though they mentioned needing evaluation on additional datasets to fully assess generalization <ref:2605.12541#pg3>. So while the results are strong on MC-MED, we need to see how robust this performs outside of that specific data distribution <ref:2605.12541#pg3>.
Rosa: That’s a fair point. So, in short, "PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation" introduces a simulator to guide latent space learning and flow matching, leading to ECGs that are not only signal faithful but also preserve critical clinical intervals <ref:2605.12541#pg3>.
Dev: It's about using mechanistic knowledge to improve the performance of purely data driven generation problems by structuring the latent space around electro-hemodynamic factors <ref:2605.12541#pg2>.
Rosa: So, to wrap up this talk on "PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation", the authors are presenting a method where they use an electro-hemodynamic simulator to co model ECG and PPG through shared cardiac phase dynamics <ref:2605.12541#pg3>.
Dev: They use this simulator to guide a Physiology Aware AutoEncoder which learns a shared latent space regularized by cross modal alignment and latency consistency from the simulator <ref:2605.12541#pg3>. And then they use that guidance in a PPG conditioned latent rectified flow to generate ECG latents while using simulator residuals to enforce ECG side morphology consistency and forward hemodynamic consistency during training <ref:2605.12541#pg3>.
Rosa: The result is a framework that combines statistical transport with physiological regularization to produce ECGs that are signal faithful, preserve critical clinical intervals, and improve cardiovascular disease classification on the large scale MC-MED dataset <ref:2605.12541#pg3>.
Dev: It’s about using simulator knowledge as soft, phase aware guidance while retaining the flexibility of data driven latent generation to bridge the gap between wearable sensing and clinical assessment <ref:2605.12541#pg3>.
Conclusion: Rosa: So, this paper is called "PPG2ECG: Physiology-Guided Latent Rectified Flow for Electro-Hemodynamic PPG-to-ECG Generation". It basically tackles how to use simple wearable pulse data, or PPG, to make an electrocardiogram because ECG is the standard for clinical cardiac assessment but it needs specialized hardware that doesn't work well for daily monitoring.
Dev: Right. They show how you can recover the electrical shape and timing from just that peripheral pulse signal, PPG, and they argue that current methods mostly rely on statistical alignment which doesn't actually structure the latent space around how physiology works <ref:2605.12541#pg1>.
Rosa: They address the problem of missing physiological awareness in the latent space abstraction. They show how you can build a simulator that models both ECG and PPG through shared cardiac phase dynamics <ref:2605.12541#pg3>.
Dev: And then they use this simulator to guide an encoder, a Physiology Aware AutoEncoder, which learns a structured latent space by enforcing consistency between the two modalities and using latency consistency derived from the simulator <ref:2605.12541#pg3>.
Rosa: Then they use that guidance in a flow matching method to generate ECG latents while using simulator residuals to enforce ECG side morphology consistency and forward hemodynamic consistency during training <ref:2605.12541#pg3>.
Dev: So, the final objective combines statistical transport with these physiological constraints to produce ECGs that are signal faithful and keep critical clinical intervals correct <ref:2605.12541#pg3>.
Rosa: It’s about using simulator knowledge as soft guidance while keeping the flexibility of data driven latent generation to bridge the gap between wearable sensing and clinical assessment <ref:2605.12541#pg3>.
Dev: The authors are showing how integrating simulator knowledge as soft, phase aware guidance can make data driven models much more robust for tasks like long term monitoring or when ECG isn't available <ref:2605.12541#pg3>.
Rosa: They also point out a limitation, though they mention they need to evaluate this on more datasets to see how it generalizes beyond the specific data they used <ref:2605.12541#pg3>.
Dev: That's fair. So, in short, this work proposes a framework that uses an electro-hemodynamic simulator to guide latent space learning and flow matching, leading to ECGs that are not only signal faithful but also preserve critical clinical intervals <ref:2605.12541#pg3>.
Rosa: It’s about using simulator knowledge as soft guidance while retaining the flexibility of data driven latent generation to bridge the gap between wearable sensing and clinical assessment <ref:2605.12541#pg3>.
Dev: Next up, we'll look at how their specific results stack up against those previous attempts at this problem.
More episodes
- Daily Summary for 2026-11-19
- 2111.07910-Mask-guided Spectral-wise Transformer for Efficient Hyperspectral Image Reconstruction
- 2203.02149-HDNet: High-resolution Dual-domain Learning for Spectral Compressive Imaging
- 2205.10102-Degradation-Aware Unfolding Half-Shuffle Transformer for Spectral Compressive Imaging
- 2311.10959-Structure-Aware Sparse-View X-ray 3D Reconstruction
- 2403.04116-Radiative Gaussian Splatting for Efficient X-ray Novel View Synthesis
- 2410.02572-Combining Pre- and Post-Demosaicking Noise Removal for RAW Video
- 2604.17453-Learned Nonlocal Feature Matching and Filtering for RAW Image Denoising
- 2603.20387-End-to-End Multi-Task Learning for Adjustable Joint Noise Reduction and Hearing Loss Compensation
- 2507.20825-Chirp-Permuted AFDM: A Versatile and Multifunctional Waveform for 6G