Two will do: Convolutional Neural Network with Asymmetric Loss and Self-Learning Label Correction for Imbalanced Multi-Label ECG Data Classification

Cristina Gallego Vázquez1, Alexander Breuss1, Oriella Gnarra1, Julian Portmann2, Giulia Da Poian1
1Sensory‐Motor Systems (SMS) Lab, Department of Health Sciences and Technology (D‐HEST), Institute of Robotics and Intelligent Systems (IRIS), ETH Zurich, 2Department of Computer Science, ETH Zurich


In this work we present a machine learning approach that is able to classify 30 cardiac abnormalities from an arbitrary number of electrocardiogram (ECG) leads. Features extracted by a deep convolutional neural network are combined with hand-crafted features (demographic, morphological, and heart rate variability metrics) and fed into a multi-layer perceptron. We employ an Asymmetric Loss(ASL) function, which enables the model to focus on hard,but under-represented, samples. To mitigate the issue of ground-truth mislabeling and to provide robustness, we investigate the use of a self-learning label correction method that iteratively estimates correct labels during training. Leaderboard results show our team SMS+1 achieved challenge scores of 0.57 0.58 0.57.56 0.57 for twelve, six, four,three, and two-lead, respectively. Our model maintains the same diagnostic potential on both standard twelve-lead ECGs and reduced-lead ECGs