Doctors size up patients the old-fashioned way: they walk in, take a look, and form a gut sense of how sick or frail someone seems.
Researchers at Mass General Brigham in Boston now want to turn that glance into a number. Their team built an AI system that studies a face photograph and estimates a cancer patient’s biological age and risk of dying. In testing, the patients the model flagged as highest risk were far more likely to die soon than the ones it flagged as lowest.
It’s a striking result, and it comes with a very large asterisk. The tools aren’t ready for clinical use, and a key question about their value is still unanswered.
And despite how it sounds, this isn’t a selfie test anyone can try at home; the research looked at people who already have cancer. So what did the team actually build, and what can it really tell us?
One Giant Face Library, Two Different Tools
The AI learned faces in general long before it learned anything about cancer.
The system, called FAHR-Face, started as a foundation model trained on more than 40 million facial images with no specific job attached. The team then adapted it into two tools, described in a study published in The Lancet Digital Health on October 19.
FAHR-FaceAge estimates biological age from a single photo, and FAHR-FaceSurvival estimates a cancer patient’s survival risk. It all follows FaceAge, an earlier Mass General Brigham system.
The motivation is surprisingly ordinary. Dr. Raymond Mak, a radiation oncologist at the hospital’s Cancer Institute, says doctors are trained to read patients at a glance: “We’re trained in medical school to walk into a room, look at the patient and make an assessment—an impression of how healthy, frail or sick they are,” he said.
He calls that impression subjective, and the team wants a more consistent version that could inform treatment discussions, spot frailty, and flag people who might need further testing. The stated aim is to support physician judgment, not replace it.
The Numbers Behind the Headline
The gap between the highest-risk and lowest-risk groups is what made this study stand out.
In one test, 64 of the 122 patients given the highest risk scores died within 90 days, or 52.5%. Among 975 patients in the lowest-risk group, about 2% died in the same window.
The 52.5% reflects one high-risk group, not the model’s overall accuracy, and it doesn’t mean every high-risk patient died.
The age tool showed a similar split. Among 1,060 patients whose estimated biological age ran more than 20 years past their actual age, 21% died within 90 days.
For the 4,195 patients estimated to look at least 10 years younger than they were, about 6% did.
Nothing here shows that looking older caused the deaths, and a younger-looking face didn’t make anyone immune.
What the AI Is Actually Looking At
Image Credit: pitinan/123RF
The two tools don’t focus on the same areas, and nobody can fully explain why the survival model picks what it does.
For survival predictions, the model’s attention landed on the bridge of the nose and the areas under the eyes, and the researchers say it’s unclear why those regions track with mortality. The age tool leaned on the nasolabial folds, forehead, temples, crow’s-feet, and cheeks.
Smokers and heavy drinkers tended to get older age estimates. Mak says that feeding both scores into a clinical prognostic model predicts survival better than either one alone.
Makeup, lighting, expression, and most cosmetic procedures didn’t throw the estimates off much. Rhinoplasty was the exception, with post-surgery photos rated slightly older.
Some of these regions echo something far older. Descriptions attributed to Hippocrates around 400 B.C. link changes in the nose, eyes, and temples to severe illness. The researchers add that the algorithms were trained on photographs and outcome or age data, not those ancient texts. Hence, the overlap is a curiosity, not evidence that the AI detects the same biology.
Related: Most Americans Drink Soda Every Day. A New Study Raises Questions About Cancer Risk
Questions That Go Beyond Accuracy
A wrong prediction could have real consequences, and the concerns extend beyond the exam room.
An independent bioethicist at a New York University center warns that a mistaken prediction could mean someone who needs care doesn’t get it. The same expert raises the possibility that insurers or employers could eventually use appearance-based health estimates.
Those are concerns, not documented cases, and nothing indicates any insurer or employer has adopted these tools.
Mass General Brigham has also filed a provisional patent application on its next-generation facial health algorithms. The team is seeking more diverse facial data and inviting the public to contribute photos for age estimation, which is separate from the cancer tool.
An outside expert calls claims about tracking aging or lifestyle changes over time premature, and there’s no reported hospital rollout, regulatory decision, or timetable yet.
Would you want your doctor to see an AI-generated score alongside their own impression of your health, and what would you want them to do with it?
More Articles You Might Be Interested in:
The post Scientists Develop AI That Can Predict Cancer Survival Risk From a Single Face Photo appeared first on FODMAP Everyday.