Authors:
Márcia Monteiro
;
Mariana Dias
and
Hugo Gamboa
Affiliation:
LIBPhys (Laboratory for Instrumentation, Biomedical Engineering and Radiation Physics), NOVA School of Science and Technology, Campus de Caparica, 2829-516, Portugal
Keyword(s):
Electrocardiogram, Signal Quality Assessment, Deep Learning, Noise, Classification, Gated Recurrent Units, Wearables.
Abstract:
The electrocardiogram (ECG) is an essential tool in the diagnosis of cardiovascular conditions. A common obstacle to readability and reliability is the vulnerability of ECG signals to noise, especially in wearable devices and long-term monitoring. Traditional methods have limited accuracy in noise detection, and, while deep learning (DL) shows promise, current models primarily focus on binary classification, lacking detailed quality analysis. This study proposes a DL model that assesses ECG signal quality, detecting and classifying specific noise types, with random-length noise segments added to clean 10-second signals to simulate real-world scenarios. The model, using gated recurrent units (GRUs), identifies three common noise types: baseline wander (BW), muscle artifacts (MA), and electrode motion (EM), achieving 98.09 % accuracy for BW, 92.62 % for MA, and 90.71 % for EM with F1 scores of 88.89 % for BW, 82.19 % for EM and 64.62 % for MA. It also surpasses existing DL methods, rea
ching 99.86 % accuracy for binary classification, with high recall and precision.
(More)