Semi-Supervised Recurrent Variational Autoencoder Approach for Visual Diagnosis of Atrial Fibrillation

In this work we propose a semi-supervised framework to visually assess the progression of time series. To this end, we present a recurrent version of the VAE to exploit the generative properties that lead it to learn in an unsupervised way a continuous compressed representation of the data. We introduce a classifier in the VAE training process to control the regulation of the latent space, allowing the network to learn latent variables that set the basis for creating an explainable evaluation of the data. We use the proposed framework to address the diagnosis of Atrial Fibrillation (AF) first validating it with simulated data with known properties and subsequently testing it with intracardiac data obtained from pacemakers and defibrillator systems.