VideoCAD: An Uncertainty-Driven Neural Network for Coronary Artery Disease Screening From Facial Videos

Coronary artery disease (CAD) is highly prevalent worldwide but is challenging to identify due to its hidden symptoms. Facial videos provide a possible means for CAD screening as facial features and skin color changes are associated with CAD risk. However, opposite test results of CAD may be drawn from facial features and skin color changes. A neural network, VideoCAD, is proposed in this work to address this problem. It consists of two modules, PulseCAD and ImageCAD. PulseCAD predicts whether CAD was present based on the facial features in a video frame, and ImageCAD performs CAD prediction based on the pulse-related features in skin color changes. VideoCAD quantifies the uncertainties of the two modules’ predictions before selecting the prediction with lower uncertainty as the final result. VideoCAD is evaluated on 1200 video samples of 200 subjects (50% CAD). Its predictions are in substantial agreement with the ground truths, with both the sensitivity (SE) and specificity (SP) of
$>0.88$
. It also outperforms the state-of-the-art methods that only consider the face or pulse features, achieving a >20% increase in Cohen’s Kappa coefficient. This technique may hold promise to achieve opportunistic screening for occult CAD in the community.