Quality Assessment of Very Long-Term ECG Recordings Using a Convolutional Neural Network

The electrocardiogram (ECG) is a physiological signal highly sensitive to disturbances during its acquisition. To palliate this issue, many works have described preprocessing algorithms operating in 12-lead, short-term ECG recordings. However, only a few methods have been introduced to detect noisy segments in single-lead, long-term ECG signals, this being a pending challenge to be resolved. Hence, this work proposes a novel technique to automatically detect low-quality segments in single-lead, long-term ECG recordings. The method is based on the high learning capability of a convolutional neural network (CNN), which has been trained with 2D images obtained when turning ECG recordings into scalograms using a continuous Wavelet transform (CWT). To validate the method, a publicly available dataset containing single-lead, long-term ECG intervals from patients with different cardiac rhythms has been used. These signals have been annotated by experts, who identified noisy intervals and those with sufficient quality to be clinically interpreted. The results have shown that the proposed method discriminates correctly between low and high-quality ECG segments with an accuracy greater than 90%, and with sensitivity slightly larger than specificity.