An efficient ECG denoising method by fusing ECA-Net and CycleGAN.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Healthcare

Background:

• Wearable electrocardiogram (ECG) acquisition is prone to motion artifacts and noise, compromising signal integrity.
• Existing denoising methods often struggle with complex noise patterns encountered in real-world wearable applications.

Purpose of the Study:

• To propose a novel end-to-end ECG denoising method for wearable devices.
• To enhance ECG signal quality by effectively removing various noise types, including motion artifacts, baseline wander, and muscle artifacts.

Main Methods:

• Developed an end-to-end ECG denoising model by fusing Efficient Channel Attention (ECA-Net) and Cycle Generative Adversarial Network (CycleGAN).
• Optimized the model using ECA-Net to emphasize crucial ECG features and a novel loss function for comprehensive feature extraction.
• Utilized ECG signals from the MIT-BIH Arrhythmia Database and various noise types from the MIT-BIH Noise Stress Test Database.

Main Results:

• The proposed method demonstrated superior denoising performance compared to existing techniques.
• Achieved significant improvements in signal-to-noise ratio (SNRimp).
• Exhibited lower root-mean-square error (RMSE) and percentage-root-mean-square difference (PRD), indicating high fidelity reconstruction.

Conclusions:

• The fused ECA-Net and CycleGAN model offers effective ECG denoising for wearable applications.
• The method shows robust generalization ability in handling diverse and mixed noise conditions.
• This approach enhances the reliability of ECG data acquired from wearable devices.