THE ARISING · Longevity & Healthspan
AI Spotted Heart Disease From Face and Tongue Photos in Small Test
A computer program that studied photos of the tongue and face separated people with coronary heart disease from healthy people with near-perfect scores in a new study. The result is striking, but it comes from a small group and carefully standardized photos. The study did not compare the method with a blood pressure cuff, a cholesterol test or a heart scan.
The work was published in Scientific Reports in 2026. It is a diagnostic study in people, not a trial of a treatment. Researchers enrolled 200 patients with coronary heart disease and 318 healthy controls, and took standardized paired photos of each person's tongue and face. Coronary heart disease is the narrowing of the arteries that feed the heart.
The team asked a practical question. Could a low-cost, non-invasive tool help screen for the disease early? In traditional Chinese medicine, the tongue is thought to reflect the heart and the face is thought to reflect blood circulation. The authors wanted to know whether combining both kinds of photo would detect disease better than either one alone.
The method used deep learning, a type of artificial intelligence that learns patterns from many labeled images. One tool outlined the tongue. Another standardized the face photos. Two image-analysis models, called ResNet50 and EfficientNet-B1, each handled tongue and face images in separate branches, and their findings were then joined.
On an independent test set, the ResNet50 model that used both photo types reached an AUC of 0.9992 and an F1 score of 0.9756. AUC measures how well a test separates sick from healthy people, where 1.0 is perfect. F1 balances missed cases against false alarms, and 1.0 is also perfect. Tongue photos alone scored an F1 of 0.9524. Face photos alone scored 0.9250.
The authors also ran 5-fold cross-validation, which repeatedly splits the data to check that results hold up. The AUCs were 0.9980 ± 0.0027 for ResNet50 and 0.9987 ± 0.0014 for EfficientNet-B1, with very little variation between splits.
The researchers also used heatmaps to show which parts of each image the models relied on. For tongues, the models focused on the central body and the coating. For faces, they focused on the area around the mouth and the cheekbone and cheek regions. The authors say this fits traditional Chinese medicine teachings. A heatmap shows where a model looked, though. It does not show that those features reflect the heart in a biological sense.
There are important limits. Only the study was available for this report, so details such as the photo equipment, how the healthy controls were confirmed to be free of heart disease, and the exact test-set size are not verified here. Scores this high in one dataset do not guarantee the same performance elsewhere. Comparing known patients with healthy volunteers is also easier than screening a mixed general public.
For you
For personal health, the takeaway is modest. It also does not compare the approach with blood pressure checks, cholesterol tests or heart scans, or suggest replacing them. Those established checks remain the proven tools.
What to watch next
Testing in other hospitals and in people with a wider mix of health conditions, which would show whether the results hold up outside the original group.
Treasury Proposes Rule That Could Restrict US Investment in Chinese Biotech Workers' Family Coverage Costs Up 9% as Premiums Outpace Wages Medicare Advantage Drug Plans Unavailable in 193 Counties for 2027, KFF Finds Treasury Poised to Propose Scrutiny of US Licensing Deals With Chinese Biotechs Alnylam's Huntington's Drug Cuts Mutant Protein About 60% in Early Study