Deep Autoencoder-Based ECG Denoising and Artifact Removal for Improved Diagnostic Accuracy

Electrocardiogram (ECG) signals are critical to identify various cardiac abnormalities. However, the accuracy of ECG-based clinical interpretations can be significantly compromised due to contamination from various artifacts. Traditional ECG denoising methods often struggle to maintain the balance between noise suppression and preservation of essential ECG characteristics. This study presents, for the first time, a deep learning (DL)-based autoencoder (AE) for robust ECG denoising and artifact cancellation and automated ECG diagnosis and classification. Unlike traditional methods, AEs are designed to learn efficient data representations, making them well-suited for distinguishing between clean signals and noise without extensive computational demands. The proposed approach leverages the PTB Diagnostic ECG Database (PTBdb) and introduces various types of noise derived from the MIT-BIH Noise Stress Test Database, including muscle artifacts (MA), electrode motion artifacts (EM), baseline wander (BW), and their combinations. The proposed AE model is trained to eliminate these noise components effectively while preserving the critical morphological features of the ECG signals. To assess the performance of the introduced method, comprehensive experiments are performed to compare clean, noisy, and AE denoised ECG signals. The AE's performance is further benchmarked against classical denoising techniques, assessing key metrics such as signal to noise ratio (SNR), mean square error (MSE), root mean square error (RMSE) and a percent root-mean square difference (PRD). The results elucidate that the introduced AE-based denoising framework not only outperforms conventional techniques in noise suppression but also enhances the reliability of ECG signal interpretation, making it an effective solution for implementation in clinical settings.