Reconstructing ECG from indirect signals: a denoising diffusion approach.

In this study, we introduce RhythmDiff, a novel diffusion-based generative model specifically designed for synthesizing high-fidelity 12-lead electrocardiogram (ECG) signals. RhythmDiff incorporates structured state-space modeling to capture morphological and temporal characteristics inherent in ECG waveforms efficiently. By embedding RhythmDiff as a prior distribution within a Bayesian inverse problem formulation, we derive the algorithm MGPS, enabling conditional ECG generation robust to varying degrees of degradations (noise, pattern of missingness) and artifacts. Our proposed framework effectively addresses the challenges associated with multi-lead reconstruction and noise reduction, demonstrating superior performance compared to existing state of-the-art ECG generative models across multiple benchmark datasets. These advancements facilitate more reliable ECG interpretation, particularly beneficial for resource-limited clinical settings and wearable technologies, enabling broader applicability in realtime cardiac health monitoring scenarios.This article is part of the theme issue 'Generative modelling meets Bayesian inference: a new paradigm for inverse problems'.