AI screen for heart disease using photos of tongue and face reaches 98% accuracy in small test
How the photo test performed
Researchers recruited 518 people: 200 with diagnosed CHD and 318 healthy volunteers. A purpose-built scanning system photographed each person’s face and tongue, and deep-learning software searched the images for patterns linked to the condition.
The system analysed face and tongue images separately before combining their results. In a test of 104 people, the best-performing version achieved 98% accuracy, identifying all 40 CHD patients while incorrectly classifying two healthy participants as having the disease.
What the AI detected
The software highlighted the tongue’s centre and coating, skin around the mouth, and cheeks. Some participants with CHD had darker tongues, thicker coatings or changes in facial pigmentation. Researchers said these areas align with traditional Chinese medicine teachings about the tongue, face and heart.
Visible signs can sometimes point to health problems: blue lips may indicate low blood oxygen, while yellow fatty deposits around the eyelids can signal very high cholesterol. But whether tongue or cheek appearance can reveal blocked heart arteries remains uncertain. The study did not establish that the identified features were caused by blocked coronary arteries.
Why screening remains uncertain
CHD occurs when fatty plaques narrow the arteries supplying the heart muscle, restricting blood flow and potentially leading to a heart attack. Doctors estimate risk using factors such as age, smoking, blood pressure, cholesterol and diabetes; suspected cases may be assessed with an ECG, blood tests or specialised scans.
A quick, painless and inexpensive photo test could help identify people who need further investigation, especially where specialist care is limited. However, this study compared people already diagnosed with CHD against healthy volunteers, rather than screening a general population with varied health conditions and cardiovascular risk.
Evidence still needed
The photographs were taken with specialised equipment under controlled lighting, and the model has not been tested in other hospitals or settings. Differences in age, other illnesses or participant characteristics could have influenced its results. Not all healthy volunteers had coronary imaging, so some may have had undetected disease.
Key points
- The best-performing model identified all 40 CHD patients in a 104-person test but produced two false positives.
- The study used controlled photographs and compared diagnosed patients with healthy volunteers, not a general screening population.
- Researchers have not shown that the facial or tongue features detected are caused by blocked arteries.
- Further studies must test performance in other settings and compare the approach with standard risk checks.