Deep Learning Classifier for Advancing Video Monitoring of Atrial Fibrillation

Abstract

Video-based non-contact monitoring of cardiac conditions offers an attractive alternative to contact-based monitoring using sensors attached to the skin. Specifically, video monitoring can significantly improve the monitoring of atrial fibrillation; a prevalent and growing cardiac disease affecting millions around the world. We propose and investigate the performance of a deep learning classifier for the detection of atrial fibrillation. We compare the performance of the proposed classifier with a benchmark of five existing classifiers based on traditional signal processing and machine learning. In addition, we compare performance across various sensing modalities, including a high-end camera, a webcam, an earlobe oximeter, and an electrocardiogram holter. To this end, we conduct a clinical study with 55 atrial fibrillation patients in a hospital setting. Results show that the proposed classifier outperforms the benchmark, especially when using a low-cost webcam, and provides consistently accurate detection when applied to an electrocardiogram, a photo plethysmography sensor, and two video camera sensors, thereby placing video monitoring on par with its contract-based counterparts.