Mamba-BiSAMNet: An ECG denoising network for multi-type noise removal

Electrocardiogram (ECG) signals are often corrupted by various types of noise, including baseline wander, muscle artifacts, and electrode motion, which can obscure waveform details relevant to clinical interpretation. These complex, time-varying interferences motivate adaptive denoising methods that can reduce noise while retaining reference-waveform structure. To address this challenge, we propose Mamba-BiSAMNet , an ECG denoising framework designed to improve waveform reconstruction from noise-contaminated signals. The model is built upon a U-Net encoder–decoder backbone enhanced with residual convolutional blocks to extract multi-scale temporal features and reinforce local feature representation. We further interpret Mamba’s S6 recurrence as an input-dependent, time-varying recursive filter that captures evolving signal structure through selective state-space dynamics. Building on this perspective, we introduce the bidirectional structure-aware module (BiSAM). This module uses backward temporal context to help assess forward features and modulates their re-integration into skip connections. By embedding BiSAM within skip pathways, the model is intended to improve temporal consistency and reduce the transmission of noise-related features. Comprehensive experiments are conducted on a synthetic dataset constructed from clean ECG signals in the QT Database and noise samples from the MIT-BIH Noise Stress Test Database. Across baseline wander, muscle artifact, electrode motion, and mixed-noise scenarios, Mamba-BiSAMNet achieves favorable reconstruction performance relative to the evaluated conventional filtering and deep learning baselines. Additionally, experiments conducted on an NVIDIA GeForce RTX 4090 show that Mamba-BiSAMNet maintains low full-window inference latency and exhibits moderate growth in peak inference memory as input length increases.