FG-DPUNet: a feature-aware gating dual-path U-Net for ECG motion artifact removal with waveform integrity

Wearable electrocardiography (ECG) has significant clinical value in ambulatory monitoring. However, motion artifacts during daily activities frequently distort faint P/T waves, compromising diagnostic interpretation. Meanwhile, single-branch or static fusion U-Net-based networks struggle to balance effective noise suppression with waveform integrity. Existing denoising methods primarily optimize global metrics such as signal-to-noise ratio improvement ( S N R i m p ), percentage root-mean-square difference (PRD), and root mean square error (RMSE), overlooking that denoising distorts P-QRS-T waveform and thus affects clinical diagnosis. This paper presents FG-DPUNet (Feature-aware Gating Dual-Path U-Net), a framework tailored to address P/T diagnostic waveform distortion induced by motion artifacts in wearable ambulatory ECG monitoring. The framework employs heterogeneous dual branches with kernel sizes of k = 7 and k = 21 to extract multi-scale ECG features in parallel, and adaptively adjusts fusion weights via a dynamic gating mechanism that operates on shared shallow features to produce temporal weights. To quantitatively assess waveform recovery quality, we introduce the P-QRS-T weighted correlation degradation index ( D w ) and P/T wave peak deviation metrics ( Δ PR , Δ R T ), providing a waveform-aware quantitative evaluation approach. On the MIT-BIH dataset under standard additive noise testing, FG-DPUNet improves the average S N R i m p by 5.94 dB and reduces the average PRD from 63.87% to 10.07% compared with state-of-the-art (SOTA) methods. On the BUT real-world motion artifact dataset, cross-subject experiments demonstrate that our model improves D w by approximately 16% over static dual-branch structures, demonstrating improved morphological preservation under motion artifacts.