3KG: Contrastive Learning of 12-Lead Electrocardiograms using\n Physiologically-Inspired Augmentations

Abstract

We propose 3KG, a physiologically-inspired contrastive learning approach that\ngenerates views using 3D augmentations of the 12-lead electrocardiogram. We\nevaluate representation quality by fine-tuning a linear layer for the\ndownstream task of 23-class diagnosis on the PhysioNet 2020 challenge training\ndata and find that 3KG achieves a $9.1\\%$ increase in mean AUC over the best\nself-supervised baseline when trained on $1\\%$ of labeled data. Our empirical\nanalysis shows that combining spatial and temporal augmentations produces the\nstrongest representations. In addition, we investigate the effect of this\nphysiologically-inspired pretraining on downstream performance on different\ndisease subgroups and find that 3KG makes the greatest gains for conduction and\nrhythm abnormalities. Our method allows for flexibility in incorporating other\nself-supervised strategies and highlights the potential for similar\nmodality-specific augmentations for other biomedical signals.\n