Robust and Task-Aware Training of Deep Residual Networks for Varying-Lead ECG Classification

Abstract

In PhysioNet/Computing in Cardiology Challenge 2021, we developed an ensemble model by combining different epochs of ResNet to classify cardiac abnormalities from 12,6,4,3,2 lead electrocardiogram (ECG) signals, where epochs are chosen based on validation performance on China Physiological Signal Challenge (CPSC) dataset and Georgia dataset. In order to adapt to the specially designed Challenge score, we designed a multi-task loss to combine the benefit of binary cross-entropy loss and Challenge loss. Besides, we also integrated a subsample frequency feature into the model to learn from the signals. To gain a better generalization ability, mixup and weighted loss are introduced. We submitted our model in the official phase with team name DataLA_NUS, and our final selected model achieved a Challenge score of 0.51, 0.51, 0.51, 0.50, 0.52 (ranked 8th, 5th, 6th, 8th, 5th) on the 12-lead, 6-lead, 4-lead, 3-lead, and 2-lead setting on the final hidden test set with the Challenge evaluation metric.