Atrial fibrillation (AF) is the most common arrhythmia and has a major impact on morbidity and mortality; however, detection of asymptomatic AF is challenging. In this work (https://www.mdpi.com/1424-8220/20/19/5517) we aim to evaluate the sensitivity and specificity of non-invasive AF detection by a medical wearable. We apply different algorithms (including a deep neural network) to five-minute periods of inter-beat intervals (IBI) for the AF detection.
A Deep neural network (DNN) is trained unsupervised on the dataset to extract relevant features for AF detection. The training objective is given by maximizing the mutual information between IBI values that are separated by a randomly chosen time point within the five-minute period. Unsupervised feature extraction followed by an unsupervised classification results in higher sensitivity and specificity compared with normalized root mean square of the successive difference (nRMSSD) an established metric for the AF detection.