MAGIC: Multi-Grained Conditional Diffusion for High-Fidelity PPG-to-ECG Translation

Building a cardiac digital twin relies on high fidelity electrocardiogram signals to accurately simulate individual cardiovascular function. However, conventional ECG acquisition is not well suited for long term, unobtrusive home monitoring, while photoplethysmography, despite its advantage in continuous collection, lacks electrophysiological information. To bridge these complementary modalities for computational analysis, we propose MAGIC, a multi-grained conditional diffusion framework for generating ECG-like waveforms from PPG. MAGIC separates conditional information into a global condition and token-wise local conditions: the former summarizes segment-level cardiovascular context and modulates the diffusion Transformer through adaptive layer normalization, while the latter preserves patch-level PPG cues and is injected through gated cross-attention. The conditional encoder is pre-trained with paired PPG and ECG signals using ROI-weighted reconstruction, contrastive learning, and latent alignment, so that PPG-derived conditions are explicitly encouraged to match ECG-derived representations. Across four public datasets and four downstream proxy tasks, MAGIC improves distributional fidelity in most settings and shows competitive task-level utility. In representative analyses, pre-training reduces FD from 0.908 to 0.693 on MIMIC-AFib; MAGIC also achieves lower FD with 50 sampling steps than RDDM with 500 steps.