Transforming ECG into a Poincaré Plot-Based Image for Its Quality Assessment through a Pre-trained Convolutional Neural Network

Wearable ECG acquisition systems present the potential to continuously monitor cardiac activity during patient's daily life.However, the acquired ECG signals are often disturbed by noise and their automated quality assessment (QA) is crucial.For this purpose, pre-trained 2-D convolutional neural networks (CNNs) have recently reported promising performance.Usually, linear timefrequency transformations have been only considered to turn the ECG signal into a 2-D image.Hence, this work aims to explore whether ECG transformation using a common approach specifically designed to visualize and feature nonlinear, complex, and chaotic dynamics in a time series, such as the Poincar plot (PP), might be helpful in QA of the signal.Then, the pre-trained AlexNet model was inputted with the image obtained by the PP-based representation of the ECG.For comparison, this network was also fed with the common wavelet ECG-based scalograms.After processes of training and testing on separate datasets, no statistically significant differences in global terms of balanced accuracy and F 1 -score were noticed for the models obtained with the two ECG-based images.However, the PP-based model was about 3 times faster than that fed with scalograms.Thus, it could be concluded that the ECG transformation into a PP-based image might be a quick and useful approach for CNN-based QA of the signal.