We all know that people often don’t look their age, which can sometimes be a good — or bad — thing.
Now some local researchers are using artificial intelligence to scan a photo of a person’s face and estimate their age based not on a birthday, but on how well that person is aging. They believe the technology could assist doctors in determining treatment options for cancer and other chronic diseases, or improve a doctor’s estimate of how long a patient might live.
“ Our study now is showing probably for the first time that we can really use AI to turn a selfie into a real biomarker source of aging,” said Dr. Hugo Aerts, Director of the Artificial Intelligence in Medicine program at Mass General Brigham, who co-authored a paper about the research.
The study, published in Lancet Digital Health, reviewed 58,851 photos of presumed healthy individuals from public data and trained an the AI program called FaceAge to assess their ages. The researchers then tested the algorithm on a group of 6,196 cancer patients using photos taken at the start of their treatment.
On average FaceAge put the cancer patients at about five years older than their chronological age, suggesting their faces indicated the toll the illness might be taking on their bodies. They also found that FaceAge outperformed clinicians in predicting short-term life expectancies of patients receiving palliative care.
The study’s co-author Dr. Ray Mak, an oncologist and faculty member in the Artificial Intelligence in Medicine program at Mass General Brigham, said doctors often make treatment decisions based on how healthy a patient appears, along with medical testing. He believes FaceAge can help doctors feel more confident when recommending types of treatment.
“We often have, as doctors, clinical intuition or gut reactions that a patient looks a lot older than they are on paper, based on the chronological age,” Mak said. “This AI-generated insight could support my clinical intuition when deciding on therapies or prioritizing tests.”
Mak recalled one of his cancer patients who looked much younger than his 86 years. Although doctors are typically hesitant to recommend aggressive cancer treatment for someone that age, the treatment was successful and the man is doing well at now 90 years old. Mak used FaceAge on the man’s photo from when treatment began and the program estimated the man’s biological age at 10 years younger than his chronological age.
He said the algorithm wouldn’t replace a doctor’s judgement or a patient’s wishes, but could serve as an additional test of a patient’s well being.
“It’s another piece of the puzzle, like vital signs, lab results, or medical imaging,” Mak said. “So we view FaceAge as a way for doctors to have more data at their fingertips to make a decision on what’s the best personalized cancer care for a given patient.”
When estimating whether a patient has six months to live, Mak said, the doctors’ accuracy was “slightly better than a coin flip.” But when using FaceAge, their accuracy improved by 20%.
Studies have shown that stress, environmental factors and illness can make someone appear older than they are. The researchers pointed to the example of the famous depression-era photo “Migrant Mother.” The woman in the photo, Florence Owens Thompson, was 32 years old when the photo was taken. FaceAge put her biological age at 46 years old.
How FaceAge determines someone’s biological age isn’t clear. The researchers theorize that the algorithm assesses muscle mass and areas around the eyes. But they said “it is a little bit like deciphering the black box algorithm” to understand if the computer is analyzing things such as hair color or wrinkles when determining a person’s biological age. They said the tool needs further study.
“I think we might be surprised at what the AI is looking at compared to what the human looks at,” Mak said. “We think of wrinkles as a sign of age, but it’s possible that the AI algorithm is not looking at those same features.”
Also under review is how FaceAge might be affected by things such as skin tone, plastic surgery and makeup. The study authors hope to open up the research to other academics to scan more photos to further test FaceAge and to establish guidelines so it’s used ethically. They say it’s not ready for clinical use, but could become a medical tool of the future.
“We look at FaceAge as just the beginning of the new field of facial health recognition AI where we can develop new algorithms to predict a variety of different health states from simple face photographs,” Mak said.
