ECG Morphological Decomposition for Automatic Rhythm Identification

Manual rhythm classification in 12-lead ECGs is time-consuming and operator-biased. We present an automatic ECG classifier using CinC's 2020 challenge dataset. In the first phase of the Challenge, 9 categories were targeted with an ensemble of 4 classifiers. In the second phase, 7 classifiers were implemented to detect 24 cardiac electrophysiological disorders. Five classifiers identified abnormalities in different specific regions of the heart's conducting system. Two classifiers were dedicated to detect premature atrial and ventricular contractions. The methodology is based on the creation of rhythm-specific intra and inter-patient templates. Firstly, signals were divided into 6 regions of interests. Secondly, for each region, intra-patient models and inter-patient rhythm-specific models were computed. The distances from each intra-patient model to each rhythm-specific inter-patient model as well as heart rate variability features and Global Electric Heterogeneity features were introduced into the classifiers. After a 10-fold cross-validation, for the provided training data in the first phase an accuracy of 94.4%±0.4, and a Challenge metric of 0.644±0.031 were obtained, whereas in the second phase an accuracy and Challenge metric of 15.0± 1.0 % and 0.030 ±0..009 were obtained.