Muscle Artifact Removal in Single-Channel Electrocardiograms using Temporal Convolutional Networks

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Medicine

Background:

• Electrocardiograms (ECG) are crucial for diagnosing heart conditions.
• Muscle artifacts (EMG) frequently contaminate ECG signals, complicating interpretation.
• Existing denoising methods often lack validation on diverse and large-scale datasets.

Purpose of the Study:

• To investigate neural network-based methods for removing muscle artifacts from single-channel ECG signals.
• To compare the efficacy of two established neural network architectures against a novel ConvTasNet-based approach.
• To evaluate method performance on larger, diverse datasets and assess out-of-distribution generalization.

Main Methods:

• Training and evaluating three neural network models: two existing architectures and a novel ConvTasNet variant.
• Utilizing simulated data from artificial mixtures of ECG (lead II) and surface EMG signals.
• Employing data from publicly available datasets: MIT-BIH Arrhythmia, PTB-XL, MIT-BIH Noise Stress Test, and TaiChi.

Main Results:

• The ConvTasNet-based method demonstrated substantial reduction of muscle artifact interference.
• The proposed ConvTasNet variant outperformed existing state-of-the-art denoising methods.
• Performance was validated on larger datasets than previously reported, including out-of-distribution testing.

Conclusions:

• The ConvTasNet variant shows significant potential for clinical application in ECG denoising.
• This advanced method can improve the reliability of ECG interpretation by reducing noise.
• The findings support the use of deep learning for enhancing diagnostic tools in healthcare.