The Conversation: An AI system trained on photographs of the face and tongue correctly identified every patient with coronary heart disease (CHD) in a small test, according to a new study from China.

Academic rigour, journalistic flair

Arts + Culture Business + Economy Education Environment Health Politics + Society Science + Tech World Podcasts Insights

Print article

An AI system trained on photographs of the face and tongue correctly identified every patient with coronary heart disease (CHD) in a small test, according to a new study from China.

The researchers recruited 518 people: 200 with diagnosed CHD and 318 healthy volunteers. Each had their face and tongue photographed with a purpose-built scanning system. The pictures were analysed using deep learning, a type of AI that learns to spot patterns in images.

The system examined the tongue and face photographs separately, then combined what each had picked up to predict who had CHD. Using both together performed better than either alone. The best-performing version was 98% accurate in a test of 104 people, correctly identifying all 40 patients with CHD, although two healthy participants were wrongly classified as having the condition – so-called “false positives”.

The AI focused on the centre of the tongue, its surface coating, the skin around the mouth and the cheeks. Some patients with CHD had darker tongues, thicker coatings or changes in facial pigmentation. The researchers say the areas highlighted by the AI are consistent with traditional Chinese medicine teachings about links between the tongue, face and heart.

Reading disease from appearance is not unique to traditional Chinese medicine. Doctors have long recognised that some visible features can point to underlying medical problems. Blue lips can signal low oxygen levels in the blood, and fatty yellow deposits around the eyelids, known as xanthelasma, can point to very high cholesterol, often caused by an inherited condition.

https://www.youtube.com/embed/y6QJceOAVY0?start=0&enablejsapi=1&origin=https%3A%2F%2Ftheconversation.com&widgetid=1&forigin=https%3A%2F%2Ftheconversation.com%2Fcan-ai-screen-for-heart-disease-using-photos-of-your-tongue-and-face-294192%3Futm_campaign%3Dthe-daily-39983%26utm_content%3Dthe-daily-39983_uk%257C2%257C39983%257Cce48b231%26utm_medium%3Demail%26utm_source%3Dbraze%26utm_term%3DCan%2BAI%2Bscreen%2Bfor%2Bheart%2Bdisease%2Busing%2Bphotos%2Bof%2Byour%2Btongue%2Band%2Bface%253F&aoriginsup=1&vf=1Coronary heart disease explained.

Whether a tongue or a cheek can reveal blocked arteries is a different matter. Skin tone, lighting, age, medication and smoking all change how a face and tongue look, and the study does not establish that these features are caused by blocked coronary arteries. AI may pick up associations that humans struggle to recognise, but that does not explain what causes them.

CHD develops when the arteries supplying the heart muscle narrow because of fatty deposits called plaques, a process known as atherosclerosis. This restricts blood flow and can eventually cause a heart attack. For many people, the first sign of a problem is chest pain, a heart attack or another life-threatening event, so spotting the disease earlier is an important goal of preventive medicine.

Doctors currently estimate risk using factors including age, smoking, blood pressure, cholesterol and diabetes. If CHD is suspected, tests can include an electrocardiogram (ECG), blood tests or specialised scans. A quick, painless, low-cost photographic test could help decide who needs those investigations, particularly in communities where access to specialist healthcare is limited.

A promising result, with questions remaining

The study, published in Scientific Reports, compared people already diagnosed with CHD against healthy volunteers, which is very different from screening the general public, where people have a mix of other illnesses and levels of cardiovascular risk. The photographs were also taken under controlled lighting with specialised equipment.

The model has not been tested on people from other hospitals or settings. AI systems sometimes perform very well on the data used to build them but less well in different hospitals, ethnic groups or healthcare settings.

Not all the healthy volunteers were assessed with coronary imaging, so undetected disease cannot be ruled out. The AI may also have been picking up differences linked to age, other illnesses or other characteristics of the participants rather than heart disease.

Screening adds a further problem. Even a highly accurate test can produce large numbers of false-positive diagnoses when it is used on a population in which the disease is relatively uncommon.

Larger studies following people over time would need to show whether this approach can pick up disease that has gone unnoticed, tell it apart from other conditions and lead to better health. Researchers would also need to compare it directly with existing ways of assessing risk. If it proves useful, it could eventually be used alongside standard heart disease risk checks, cholesterol tests and blood pressure measurements.