A Rectified Flow-Based Unified Framework for ECG and PPG Denoising
Abstract
Artifacts and noise are commonly encountered during the acquisition of electrocardiogram (ECG) and photoplethysmogram (PPG) signals, which can degrade signal quality and compromise the reliability of downstream analysis and decision-making systems. To address this issue, we propose a rectified flow-based denoising framework integrated with a residual self-attention (RSA) mechanism, enabling effective modeling of complex temporal dependencies and noise characteristics. The proposed approach is systematically evaluated on both ECG and PPG datasets. The experimental results demonstrate that the model consistently outperforms state-of-the-art (SOTA) denoising methods across multiple evaluation metrics, including waveform fidelity, morphological preservation, and downstream task performance. Furthermore, we assess the computational efficiency of the proposed framework by analyzing its inference latency on an edge device. In addition, the proposed method is validated on real-time ECG recordings acquired using a BioRadio device during different subject activities, demonstrating its ability to effectively suppress naturally occurring noise and motion artifacts while preserving the underlying ECG morphology.