A stacked contractive denoising auto-encoder for ECG signal denoising

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Healthcare

Background:

• Electrocardiogram (ECG) signals are crucial for diagnosing cardiac diseases.
• ECG signals are frequently corrupted by noise, including baseline wander, electrode contact noise, and motion artifacts, hindering accurate diagnosis.
• Existing denoising auto-encoders (DAEs) require improvement for robust ECG signal processing.

Purpose of the Study:

• To introduce a novel contractive denoising technique to enhance ECG signal denoising.
• To develop a stacked contractive denoising auto-encoder (CDAE) for multi-level feature extraction and noise reduction in ECG signals.
• To evaluate the performance of the proposed CDAE against conventional ECG denoising methods.

Main Methods:

• Development of a stacked contractive denoising auto-encoder (CDAE) utilizing a deep neural network (DNN).
• Incorporation of a contractive penalty based on the Frobenius norm of the Jacobian matrix for learned features.
• Evaluation using ECG signals from the MIT-BIH Arrhythmia Database and noise from the MIT-BIH Noise Stress Test database.

Main Results:

• The proposed CDAE algorithm demonstrates superior performance compared to conventional ECG denoising methods.
• Achieved an improvement of over 2.40 dB in signal-to-noise ratio (SNR).
• Showed significant reduction in root mean square error (RMSE), ranging from 0.075 to 0.350.

Conclusions:

• The developed CDAE is an effective deep learning approach for denoising ECG signals.
• The method enhances ECG signal expression through multi-level feature extraction, leading to improved diagnostic accuracy.
• CDAE offers a promising solution for improving the reliability of ECG-based cardiac diagnostics.