CausalDiffusion: A Causality-Embedded Diffusion Model for Cross-Modal Physiological Signal Synthesis

ICLR 2026 Conference Submission17328 Authors

Keywords: Denoising Diffusion Probabilistic Models; Causal Inference; Dynamic Causal Graph; Physiological Signal Synthesis

Abstract: Synthesizing high-fidelity, physiologically plausible signals offers transformative solutions to data scarcity and privacy concerns in biomedical research and clinical care. While Denoising Diffusion Probabilistic Models (DDPM) excel at producing statistically plausible signals, they often fail physiological validation due to neglecting the underlying causal mechanisms governing cardiovascular dynamics. To bridge this gap, we introduce CausalDiffusion, a novel Causality-Embedded DDPM that learns and embeds dynamic causal knowledge directly into the reverse diffusion process. One core of CausalDiffusion is to learn sample-specific causal graphs by dynamically weighting a static base graph, which is constructed offline by synergizing domain knowledge with causal structures discovered by the Fast Causal Inference (FCI) algorithm. These adaptive causal embeddings modulate the U-Net denoiser at multiple scales, transforming causal graphs into structured regularizers that constrain the generative pathways to physiologically plausible manifolds. On the challenging task of continuous blood pressure waveform generation from ECG and PPG inputs, CausalDiffusion achieves state-of-the-art performance, demonstrating remarkable robustness, particularly under arrhythmic conditions where pure data-driven models degrade. Furthermore, we showcase the framework's generality by applying it in the PPG-to-ECG synthesis task, where it achieves superior performance in generating complex waveform morphology. By synergizing the generative power of diffusion models with the rigor of causal reasoning, our work establishes a new paradigm for building reliable and interpretable generative models in biomedicine and beyond.

Primary Area: generative models

Submission Number: 17328