[2Yin-A] Poster 3

[2Yin-A-27] Long-Context Modeling and Physiological Constraints for ECG Generation from PPG

〇So Kariyama^1,3, Kazuyuki Yoshikawa^2,3, Mika Akama³, Nayuta Mizuguchi^1,3 (1. The University of Tokyo, 2. University of Hyogo, 3. FastNeura Inc.)

Keywords:

Physiological Signal Generation,Conditional Generative Models,Electrocardiogram (ECG),Photoplethysmogram (PPG)

Generating electrocardiograms (ECGs) from wearable photoplethysmograms (PPGs) can enable low-cost and continuous cardiac monitoring, but existing PPG-to-ECG generators often reproduce waveform shapes while failing to preserve beat timing. We propose a two-stage RR-conditioned Rectified Flow framework for long-context ECG generation from PPG. In Stage A, a lightweight CNN predicts an RR-interval signal from PPG using differentiable rhythm-consistency losses. In Stage B, the predicted RR signal is frozen and provided as an additional condition to a Rectified Flow generator with a 1D U-Net backbone, enabling explicit control of beat timing without relying on non-differentiable peak-based losses during training. Experiments on the PPG-DaLiA dataset with 30 s and 60 s windows show that the proposed method improves HR MAE by 55-67% and RR MAE by 14-30% over a baseline, while maintaining waveform RMSE and PSD similarity. HRV improvements remain limited, indicating the need for better R-wave morphology modeling.