Machine Learning Approach to Detection of Atrial Fibrillation Using High Quality Facial Videos

Videoplethysmography (VPG) is an emerging technology that uses a video camera to capture subtle skin color variations caused by blood volume flow. VPG offers a seamless, non invasive long term cardiac monitoring solution which is necessary to capture a-symptomatic heart conditions such as early stages of Atrial Fibrillation (AF). In this work we investigate the ability of VPG based on high end Basler RGB camera to identify AF in a controlled hospital setting. We conduct a clinical study and explore various modern Machine Learning (ML) algorithms and feature extraction methods to provide classification among three groups: healthy subjects, subjects with AF before and after they undergo direct current cardioversion. Our results reveal that when features represent statistical, dimensional and time-frequency properties of the underlying VPG signal, all non-linear and ensemble classifiers under test achieve close to perfect performance in both 2- and 3-class classification. Among them, random forest and extreme gradient boosting classifier consistently attain the highest accuracy, sensitivity and specificity of almost 100%. These results suggest that the VPG technology relying on a high quality camera combined with intelligent ML classifiers and smart feature selection can be used to facilitate diagnosis of heart conditions within a controlled hospital environment and has a great potential to go beyond.