Cordelia: An Application for Automatic ECG Diagnostics

Abstract

The authors present a prototype of an application named Cordelia, which enables the prediction of selected cardiac findings on standard 12-lead ECG recordings. The application is based on an ensemble model consisting of ten deep residual convolutional neural networks. In order to eliminate the different scope of the assessed labels, as well as the different approach in assessing the presence (or absence) of certain labels in different datasets, the model was trained using 3-valued logic. Cordelia allows not only to determine the probability value of each of the assessed labels, but also to draw an ECG recording and evaluate the technical conditions of the record, which can have negative impact on the prediction outcomes (e.g., significant baseline shift, signal outages, etc.) The application can be beneficial especially for primary care physicians less experienced in the evaluation of ECG recordings. As a part of the telemedicine platform, it could enable very fast consultation of practitioners with specialists without the need for a physical visit of patient. The basis of the developed solution can also be used to create models for evaluating the presence of arrhythmia in long-term ECG recordings (Holter monitoring) with reference to the location and duration of the episode(s).