AI trained on 40 million face images can estimate biological age and cancer survival odds from just one photo

Researchers at Mass General Brigham in Boston have built AI models that estimate a cancer patient's biological age and risk of death from a single photo of their face.

Of 122 patients the survival model rated highest risk, 52.5% died within 90 days of the assessment, compared with about 2% of the 975 it rated lowest risk, according to the study, published Oct. 19 in The Lancet Digital Health.

"I was pretty surprised to find out that we could pick up signs of mortality risk from just a facial photograph," study author Raymond Mak, a radiation oncologist at Mass General Brigham Cancer Institute, told CNN.

The team first built a foundation model called FAHR-Face, trained on more than 40 million facial images spanning a range of ages and ethnic backgrounds.

From it, they developed two tools. FAHR-FaceAge estimates biological age, a measure of how the body has aged rather than how many years a person has lived. FAHR-FaceSurvival predicts survival in cancer patients directly from a face photo.

Among the 1,060 patients whose estimated biological age exceeded their chronological age by more than 20 years, 21% died within 90 days. About 6% of the 4,195 patients estimated to be at least 10 years younger than their actual age died in the same period.

The two tools read different parts of the face. The age model focuses on the temples, the forehead above the eyes and the nasolabial folds, the lines running from the nose to the corners of the mouth, while the survival model looks at the bridge of the nose and the area under the eyes.

Those regions overlap with the signs the Greek physician Hippocrates described in the Book of Prognostics around 400 B.C., when he listed "the nose sharp, the eyes sunken, the temples fallen in" as signs of muscle wasting and mortality risk.

"To think that our algorithms might be replicating that wisdom of Hippocrates from millennia ago was mind-blowing to us," Mak said. The models had no access to ancient texts, he noted, as they were trained "purely on face photographs and then whether somebody died or not, or their age."

The AI’s survival model (L) and biological-age model (R) weigh different parts of the face, with yellow marking the areas that count most. Photo courtesy of Mass General Brigham

According to Mass General Brigham, the age model's estimates held steady across changes in pose, lighting, facial expression and makeup, and across most plastic surgery procedures tested, shifting slightly only after rhinoplasty.

The authors say the tools are not meant to replace doctors. Mak said catching accelerated aging or high mortality scores early could help identify patients who may benefit from further testing for serious conditions, and that more research could show whether the findings extend to healthy people.

Wilbur Lam, a professor of pediatrics and biomedical engineering at Emory University who was not involved in the study, called it "an extremely compelling study that stems from a large data set." "Its true clinical utility, however, will need to be demonstrated prospectively," he told CNN.

Miroslaw Bober, a professor of video processing and machine intelligence at the University of Surrey in the U.K., said the study provides "convincing evidence" of a statistical link between facial appearance and survival in cancer patients receiving radiotherapy.

But "predicting an outcome is not the same as explaining it," he said in a statement distributed by the Science Media Center, adding that the study did not test whether the AI recognizes anything a doctor would not. "This is a promising proof of concept, but it is not ready for clinical use."

Emma Pierson, an assistant professor of computer science at the University of California, Berkeley, raised concerns about racial bias. "Any time you train something on someone's face, you are implicitly including patient race in the model's predictions," she told The Washington Post.

The researchers noted that their cohorts included only cancer patients, mostly from high-income countries, and called for validation in broader, more representative populations.

Mass General Brigham has also opened an online FaceAge research study, in which members of the public can upload a face photo, receive a research readout of their facial age and contribute data to help validate the tool.