SkipDAEformer: A High-Precision Representation Learning Method for Removing Random Mixed Noise in MCG Signals

Area of Science:

• Biomedical Engineering
• Signal Processing
• Cardiology

Background:

• Automated analysis of magnetocardiography (MCG) is crucial for cardiovascular disease diagnosis and prediction.
• Clinical MCG signals frequently contain noise, hindering accurate analysis.
• Existing denoising methods struggle with long-term MCG signals and complex spatial structures.

Purpose of the Study:

• To introduce a novel, high-precision, robust representation learning method (SkipDAEformer) for denoising MCG signals.
• To enhance the extraction and fusion of temporal and spatial information for improved MCG signal analysis.
• To develop a method capable of effectively separating clean MCG signals from random mixed noise.

Main Methods:

• Proposed SkipDAEformer, a denoising autoencoder integrating attention fusion mechanisms.
• Employed skip connection multi-scale feature fusion to capture long-range dependencies and spatial features.
• Utilized global feature fusion to refine semantic information and learn comprehensive MCG signal representations.

Main Results:

• SkipDAEformer demonstrated superior denoising performance, channel consistency, and feature consistency compared to existing methods.
• The method exhibited strong generalization ability and can be adapted to a self-supervised learning framework.
• Experimental results confirmed SkipDAEformer's effectiveness in noise reduction and diagnostic classification tasks.

Conclusions:

• SkipDAEformer offers a significant advancement in denoising MCG signals, outperforming traditional techniques.
• The method shows high clinical acceptability and diagnostic value, potentially improving cardiovascular disease assessment.
• This approach advances automated analysis of MCG data, benefiting clinical practice and research.