Mobile Menu

Say cheese: Deep learning detects coronary artery disease based on photos

A team of researchers have developed and validated a deep learning algorithm for detecting coronary artery disease (CAD) based on facial photos.

Coronary artery disease

CAD remains the leading cause of death and chronic disability in cardiovascular diseases across the world. As a result, there is an urgent need for precise, practical and cost-effective screening tools. Aside from clinical risk factors, researchers have found that some facial features are associated with increased risk of CAD. These include alopecia, grey hair, facial wrinkles, earlobe creases, xanthelasmata and arcus corneae. These features increase risk of CAD and poor cardiovascular health. The use of facial features for CAD screening have been limited by: (1) few categories and low prevalence of facial features, (2) lack of specific definitions, and (3) poor reproducibility in human identification.

As artificial intelligence (AI) has evolved, deep learning algorithms have become a promising tool for diagnosis and prognosis of disease. Implementation of AI technologies in clinical practice have already started, including interpretation of images, pathology slides and electrocardiograms. In this study, published in the European Heart Journal, researchers hypothesised that deep learning may help in interpreting facial features from photos to detect CAD.

A deep learning algorithm

The team conducted a multicentre cross-sectional study. Researchers obtained data from two studies at nine sites in China. The team first enrolled 5,796 patients, who they then randomly divided into a training group and a validation group for algorithm development. They also enrolled another 1,013 patients for inclusion in the test group for algorithm testing. Patients who were undergoing coronary angiography or computed tomography were eligible. Researchers used deep convolutional neural networks to train an algorithm to detect CAD from facial photos.  

They found that the CAD detection algorithm had an overall sensitivity of 0.80 and specificity of 0.54 in the test group. The Area Under the Curve (AUC) was 0.730. The algorithm also outperformed the Diamond-Forrester model and the widely-used CAD consortium clinical score.

This study provides further evidence for the feasibility of using human facial features to detect CAD. It also supports the use of deep learning tools for assessing and screening patients in the clinic or community. The researchers believe that this will help guide further diagnostic testing and medical visits.

Image credit: Background vector created by freepik –

Share this article