A dynamic transfer network for cross-database atrial fibrillation detection

Deep learning has been successfully applied to the automatic diagnosis of cardiovascular disease. However, the domain shift of electrocardiogram (ECG) signals are huge since the ECG signals are generated in different acquisition environments, such that the model trained on a specific dataset typically performs worse when directly applied to a new dataset. In this paper, a dynamic transfer network (DTN) is proposed and applied to the cross-database AF detection. We devise a more general two-phase domain adaptation framework. Firstly, in the pre-training phase, three convolutional neural networks are trained as pre-trained models using the ECG signals from the source domain. Secondly, in the domain adaptation phase, a dynamic adaptive module (DAM) is introduced to mitigate the impact of the distribution differences by adaptively learning the ECG features of source and target domains. Furthermore, the minimum class confusion (MCC) loss is used to enhance the class discriminability to achieve highly accurate AF detection on the target domain. We performed six transfer tasks on three public ECG databases: the MIT-BIH Atrial Fibrillation Database (AFDB), the 2017 PhysioNet/CinC Challenge Database (Phy2017), and the China Physiological Signal Challenge 2018 Database (CPSC2018). The DTN-AlexNet obtained 91.35% accuracy and 87.92% F1 score on the transfer task AFDB-*Phy2017. The DTN-VGG11 obtained 93.38% accuracy and 89.50% F1 score on the transfer task AFDB-*CPSC2018. The DTN-ResNet obtained 97.58% accuracy and 96.83% F1 score on the transfer task Phy2017-*AFDB. The experimental results demonstrate that the proposed DTN reduces the impact of the distribution differences, and performs well on cross-database AF detection.