Integrated Res2Net combined with Seesaw loss for Long-Tailed PCG signal classification

Abstract

PCG signal contains important information about heart movement, which is of great significance to the diagnosis and prevention of heart disease. In this paper, we adopt Res2Net which is a multi-scale neural network as the backbone framework to train on PCG dataset. Meanwhile, to address the problem of data imbalance, we utilize Seesaw loss to replace the traditional Cross-entropy loss. Seesaw loss uses mitigation factor and compensation factor to re-balance the gradient of positive and negative samples to reduce the dominance of head classes in the training process. Moreover, we propose an integrated method which is to select three models with the best performance on the test set to integrate to improve the generalizability of Res2Net and the accuracy of PCG classification. Furthermore, we conduct extensive experiments on PCG datasets, and the results show that our method is effective and has strong competitiveness.