Automatic Attention Learning Using Neural Architecture Search for Detection of Cardiac Abnormality in 12-Lead ECG

In this article, a new attention learning method based on neural architecture search (NAS) is introduced to detect cardiovascular diseases (CVDs). Cardiac abnormalities are a series of CVDs that threaten human health. Electrocardiography (ECG) is a useful tool to express the status of cardiac activity and detect cardiac abnormalities. In recent years, the convolutional neural network and attention mechanism have been applied in ECG classification with good feature-learning ability. However, there are a large number of hyperparameters that must be manually adjusted, so that they have limited ability to learn temporal and channel information from feature maps in the network. Therefore, in this article, a new attention learning method based on NAS is proposed, so that the network can better learn the temporal and channel information in the ECG and perform better than artificially designed architectures. The results were verified using the 12-lead ECG dataset of the China Physiological Signal Challenge (CPSC) 2018, which contains eight different CVDs and the proposed method performs better than state-of-the-art methods.