A Deep Learning Solution for Automatized Interpretation of 12-Lead ECGs

Abstract

A broad variety of algorithms for detection and classification of rhythm and morphology abnormalities in ECG recordings have been proposed in the last years. Although some of them have reported very promising results, they have been mostly validated on short and non-public datasets, thus making their comparison extremely difficult. PhysioNet/CinC Challenge 2020 provides an interesting opportunity to compare these and other algorithms on a wide set of ECG recordings. The present model was created by 'ELBIT' team. The algorithm is based on deep learning, and the segmentation of all beats in the 12-lead ECG recording, generating a new signal for each one by concatenating sequentially the information found in each lead. The resulting signal is then transformed into a 2-D image through a continuous Wavelet transform and inputted to a convolutional neural network. According to the competition guidelines, classification results were evaluated in terms of a class-weighted F-score (Fßand a generalization of the Jaccard measure (Gß). In average for all training signals, these metrics were 0.933 and 0.811, respectively. Regarding validation on the testing set from the first phase of the challenge, mean values for both performance indices were 0.654 and 0.372, respectively.