Multi-label Disease Diagnosis Based on Unbalanced ECG Data

Abstract

With the improvement of living standards, heart diseases have become one of the common diseases that threaten health of human beings. Electrocardiogram (ECG) is an important basis for diagnosing heart diseases. In this paper, we propose a model to predict 55 classes of heart diseases simultaneously, that is, to solve a multi-label classification task. In order to make full use of the characteristics of the ECG, we propose a network structure combining residual neural network (ResNet) and gated recurrent unit neural network (GRU). On this basis, in order to solve the problem of imbalanced data set, the loss function is a improved focal loss. The results of experiments show the effectiveness of our method. More specifically, the method improves F1 score, while the hamming loss is reduced. Observing the classify result of each single class, we improve F1 score and average area under the receiver operating characteristic curve (AUC) for most classes.