A Two-Phase Multilabel ECG Classification Using One-Dimensional Convolutional Neural Network and Modified Labels

Peter Bugata1, Peter Bugata Jr.1, Vladimira Kmecova1, Monika Stankova1, David Gajdos1, David Hudak1, Richard Stana2, Simon Horvat2, Lubomir Antoni2, Gabriela Vozarikova2, Erik Bruoth2, Alexander Szabari2
1VSL Software, a.s., 2Pavol Jozef Šafárik University


Within PhysioNet/Computing in Cardiology Challenge 2021, we developed a two-phase method of automatic ECG recording classification. In the first phase, we pre-trained a model on a large training set with our proposed mapping of original labels to the SNOMED codes, using three-valued labels. To solve the multilabel binary classification task, we used a deep convolutional neural network, which is a 1D variant of the popular ResNet50 network. In the second phase, we performed fine-tuning for the Challenge metric and conditions. In the official round, our team CeZIS obtained the Challenge metric score of 0.717, 0.680, 0.703, 0.702, and 0.681 on the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead validation datasets, respectively.