Physiology-Informed Diffusion for 12-Lead ECG Generation

Abstract:

Large-scale 12-lead ECG data are critical for training reliable cardiac machine learning
systems, yet their availability is limited by privacy constraints, annotation cost, and severe
class imbalance. Generative models offer a promising solution, but standard diffusion models
typically treat ECGs as generic multivariate time series and do not explicitly exploit known
physiological structure.
We propose PhysDiff-ECG, a physiology-guided diffusion framework that integrates cardiac
ordinary differential equation (ODE) prior into the diffusion trajectory. Our central idea
is to make ECG physiology tractable during training by deriving differentiable regularizers
from a dynamical model of cardiac activity together with a differentiable 12-lead observation
model. Given a denoised reconstruction along the reverse process, PhysDiff-ECG fits a latent
physiological explanation via an unrolled inner optimization and penalizes violations of both
the simulator dynamics and the induced ECG reconstruction.
This training-time regularization biases the learned denoising trajectories toward physio-
logically realizable ECGs while preserving the flexibility of latent diffusion. Experiments
on standard 12-lead ECG benchmarks show that PhysDiff-ECG improves physiological fi-
delity, representation-space realism, class-conditional diagnostic consistency, and downstream
classification performance relative to strong GAN and diffusion baselines.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Daniel_Durstewitz1

Submission Number: 8564