Transfer Learning for Electrocardiogram Classification Under Small Dataset

摘要

The First China ECG Intelligent Competition is held by Tsinghua University. It is aimed to intelligently classify electrocardiogram (ECG) signals into two categories in preliminary and nine categories in rematch. The detailed ECG categories are listed in subsequent section. Our team proposes a deep residual network for diagnosing cardiovascular diseases automatically based on ECG, making full use of the network's hierarchical feature learning and feature representation ability. Considering that the amount of this competition data is small, especially in the stage of preliminary where there are only 600 training samples, while the deep learning-based method is data-hungry. Transfer learning idea is introduced into the training process of proposed deep neural networks. The proposed network is firstly trained on the Physionet/CinC Challenge 2017 dataset that is an open-public ECG data with single lead. Then it is continuously fine-tuned on the competition dataset with 12 leads. The performance of the proposed network is improved a lot. The proposed method achieves F_1 score of 0.89 and 0.86 in the hidden test set of preliminary and rematch, respectively. The research code will be released later.