AriReGen: An Autoregression-Induced Residual Generative Approach for Robust ECG Denoising

To facilitate the analysis of electrocardiogram (ECG) signals collected in challenging environments, denoising algorithms must be adaptable to diverse noise sources and varying intensities. This study aims to improve the generalizability of denoising models trained on specific datasets, allowing effective performance in previously unseen scenarios. We propose a two-stage autoregression-induced residual generative (AriReGen) approach. In the first stage, an autoregression model extracts the intrinsic temporal structures of the ECG signal. In the second stage, AriReGen operates on the input-output residuals via flow matching. The first stage requires no labeled data for training. The second-stage conditional residual generation simplifies the domain transform task under the temporospectral guidance of the autoregression output. We established a comprehensive evaluation protocol covering waveform fidelity, downstream applicability, and clinical value. Extensive open-set experiments across multiple public datasets demonstrate that AriReGen outperforms state-of-the-art baseline methods in denoising accuracy, while exhibiting stable generalization and enhanced interpretability. As it helps a denoised ECG provide diagnostically reliable insights, AriReGen shows strong potential for facilitating accurate ECG examination in real-world clinical applications.