Can AI spot heart disease in a face or tongue photo? A small study is only a first step

A photo-based AI system identified all 40 people with coronary heart disease in a small test, but that is not the same as proving it can screen the public. The study’s size and controlled conditions leave a long road from promising result to useful medical test.

Watch Desk analysis

What happened

The Conversation describes a study in which researchers used a purpose-built scanner to photograph 518 people: 200 already diagnosed with coronary heart disease and 318 healthy volunteers. A deep-learning system analysed face and tongue images, separately and together. The best-performing version was reported as 98% accurate on a test of 104 people, identifying all 40 participants with the disease while incorrectly flagging two healthy participants.

The report says the researchers highlighted features including the centre of the tongue, its coating and skin around the mouth and cheeks. The study was published in Scientific Reports. The Conversation notes that the photos were taken with specialised equipment under controlled lighting, and that the system has not been tested across other hospitals or settings.

Key findings

• All 40 diagnosed cases were identified in the test
  The same model also produced two false positives among healthy participants.
• The test compared diagnosed patients with healthy volunteers
  That is not the same population mix a real-world screening programme would encounter.
• The photos used controlled conditions and specialised equipment
  Performance elsewhere, including across different settings and groups, remains unknown.

Why it matters

A quick, painless photo test could be useful if it helps clinicians decide who needs established investigations, particularly where access to specialist care is limited. But a system that separates people already known to have the disease from healthy volunteers has not yet shown it can reliably find previously undetected cases in the general population.

False positives matter too. When a disease is less common in the people being screened, even a test that performs well in a small study can flag many people who do not have it. The report also notes that age, other illnesses and participant characteristics could explain some of the patterns the AI detected; the study does not establish that the visible features are caused by heart disease.

Our read

This is an interesting research signal, not a reason to diagnose anyone from a selfie. The useful next steps are larger studies over time, tests in different settings and a direct comparison with current risk assessments. Until then, the headline number deserves interest, not a place in the clinic waiting room.

What to watch

• Whether larger studies test people whose heart-disease status is not already known.
• Whether the system works across hospitals, lighting conditions and different groups of people.
• Whether it adds useful information beyond established risk checks and tests.

Discussion spark: What evidence should a photo-based AI test need before doctors use it to decide who gets further heart investigations?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.