A Snapshot of a Face Turned Into a Cancer Test
A single photo of a cancer patient's face, fed to an algorithm, yielded a score that tracked who was likely to die. Among the 122 patients it flagged as highest risk, 64 died within 90 days of the picture.
Those 122 came from 3,820 people with cancer used to test the model in a study that built and validated it. Lancet Digital Health carried the work, from a Mass General Brigham team led by radiation oncologist Ray Mak. A second check, on more than 38,000 people with cancer, found that those whose faces read older than their actual age died at higher rates.
The algorithm first absorbed 40 million face images with no assignment. The team then tuned it to estimate biological age and to gauge survival odds. Used together, the two scores predicted survival better than either did alone, Mak says.
Next the researchers mapped where the software was looking. For age, it fixed on the creases running from nose to mouth and on patches of forehead. For death risk, it settled on the nasal bridge and the skin under the eyes. Why those spots carry the signal is unclear.
Part of it may be familiar risks. Smokers and heavy drinkers tended to look older to the AI. The tool isn't ready for clinics. Its survival model learned from a mostly white Boston-area patient group, and NYU bioethicist Matthew Liao cautioned that a wrong call could steer care away from someone who needs it.
Smiles, frowns, makeup and lighting didn't throw it off. A nose job did. Patients tended to look slightly older to the AI after the surgery.
Read the full story at Scientific American, October 19, 2026
Hot Take: Mirror, mirror, on the wall, what can faces tell us all? The wicked queen's mirror has gone digital, and it's developed a disturbing interest in noses.
Subscribe to our newsletter.
Be the first to know - subscribe today
Member discussion