Elimination of Random Mixed Noise in ECG Using Convolutional Denoising Autoencoder With Transformer Encoder

Area of Science:

• Biomedical Signal Processing
• Artificial Intelligence in Healthcare
• Cardiology

Background:

• Electrocardiogram (ECG) signals are susceptible to various noise types, including baseline wander (BW), electrode motion (EM) artifacts, and muscle artifact (MA).
• Combined noise often leads to misinterpretations by cardiologists, hindering accurate diagnosis.
• Existing denoising methods may introduce distortions or fail to handle complex, mixed noise effectively.

Purpose of the Study:

• To develop a novel deep learning model for suppressing random mixed noise (RMN) in ECG signals with minimal distortion.
• To improve the accuracy and reliability of ECG interpretation in the presence of noise.
• To enhance the performance of R peak detection in denoised ECG signals.

Main Methods:

• A Transformer-based Convolutional Denoising AutoEncoder (TCDAE) model was proposed, featuring stacked gated convolutional layers and a Transformer encoder with multi-head self-attention.
• A frequency weighted Huber loss function was introduced during training to minimize distortion in both time and frequency domains.
• The TCDAE model was trained and validated on the QT Database (QTDB) and MIT-BIH Noise Stress Test Database (NSTDB), with generalization tested on the Icentia11k database.

Main Results:

• The TCDAE model demonstrated robust performance in removing RMN and individual noise types compared to baseline methods across QTDB and NSTDB.
• On the Icentia11k database, TCDAE achieved state-of-the-art results, reducing false positives in R peak detection by 55% post-denoising.
• The model effectively captured both short-term and long-term ECG signal characteristics, maintaining stability under severe noise conditions.

Conclusions:

• The proposed TCDAE model offers a highly effective solution for denoising ECG signals corrupted by random mixed noise.
• TCDAE exhibits superior performance and generalization capabilities compared to existing methods, significantly improving R peak detection accuracy.
• The model's computational efficiency makes it suitable for clinical application, promising enhanced diagnostic reliability in real-world scenarios.