DAAT: A New Method to Train Convolutional Neural Network on Atrial Fibrillation Detection.

Atrial fibrillation (AF) is a common disease in elderly people which is associated with high morbidity. Detecting AF with electrocardiogram (ECG) recordings benefits them for early diagnose and treatment. Lots of models based on convolutional neural network (CNN) have been proposed for such purpose. However, how to train such models so as to get better performance still remains challenging. In this paper, we put forward a dynamic attention assistant training (DAAT) process for CNN model training, which can not only improve the accuracy of verified strong ResNet on AF detection task, but also help hardly trained DenseNet to get a good performance under the precondition of a low proportion of positive AF samples, which usually occurs in real tasks. The training process works even when some attention layers have already been utilized within convolutional layers like SENet. (The source code can be downloaded from https://github.com/mszjaas/DAAT).