Adaptive Latent Space Domain Transfer for Atrial Fibrillation Detection

Abstract

Atrial fibrillation (AF) is characterised by disorganised atrial electrical activity and contraction. The complications of atrial fibrillation is varied, so it shows several modalities in the Electrocardiogram (ECG). According to our statistics, the beat features and duration time of AF are different from person to person. The distribution of available annotated source ECG date is not as the target. Due to the rapid development of portable monitor, the ECG data explodes. The existing algorithms for the detection of AF perform well only in the training database. Training a general used model requires a great deal of labeled data for every user, this is a huge amount of work that is almost impossible. In order to make full use of the limited labeled ECG data and training a general model, we propose an adaptive latent space domain transfer method for the detection of AF. The model learned from the source data is automatically adapted with little annotated or none in the target data. The MIT-BIH Atrial Fibrillation Database is regard as standard reference for classifier. Then we carry the experimental verification on the MIT-BIH Normal Sinus Rhythm Database and the Long-Term AF Database. The transfer method shows better performance than directly applied. It does make sense for detection and analysis of clinical dynamic electrocardiogram and individual ECG monitoring.