A dual-path interactive denoising autoencoder for removing mixed noise in multi-lead electrocardiogram signals

Abstract

Denoising of dynamic electrocardiography (ECG) is a hotspot in ECG processing. Existing denoising methods mainly target single-lead signals or single noise sources, with limited research addressing mixed noise across multiple leads. Based on the time-frequency characteristics of common noise in ECG acquisition, this study introduces the Dual-Path Interactive Denoising Autoencoder (DP-IDAE). One pathway utilizes multi-scale convolutional kernels to extract local features, while the other employs bidirectional long short-term memory to capture global features. An interactive transmission mechanism facilitates information exchange between pathways. Experimental results demonstrate DP-IDAE's superior denoising performance across seven types of noise, including single and mixed noise composed of baseline wander (BW), muscle artifact (MA), and electrode motion (EM). In the complex noise environment of BW+EM+MA at -6 dB, the$\text{SN}{R}_{\text{imp}}$and PRD still achieve 8.3 dB and 31.87 %, respectively. Additionally, the study investigates the relationship between denoising effects and signal similarity of different lead ECGs. It concludes that a higher similarity between lead signals leads to better denoising performance, especially for leads I, II, AVF, and V4-V6, where the denoising effect is more significant. From both time-domain and frequency-domain perspectives, DP-IDAE effectively removes local and global noise by leveraging the advantages of dual pathways.