A Face Photo Now Carries a Cancer Survival Score

Mass General Brigham researchers published FAHR-Face on October 19, 2026, an AI that scores cancer survival from one face photograph. The Boston health system trained the foundation model on more than 40 million facial images, then adapted it into an age estimator and a mortality ranker for people receiving radiation for cancer.

Doctors already judge frailty by looking. This work turns that look into a number that can sit in a chart, travel to a public website, and follow a patent filing, while the team still cannot say which facial features drive the prediction.

Two Scores From One Photograph

FAHR-Face is short for Foundation Artificial Intelligence for Health Recognition-Face. Fridolin Haugg and Grace Lee, joint first authors in the Artificial Intelligence in Medicine program at Mass General Brigham, built it as a masked autoencoder on the WebFace42M photo set, then fine-tuned two heads for different jobs.

Hugo Aerts, director of that program and co-senior author, said the shared backbone is the technical step. “The key technical advance is that FAHR-Face learns a general representation of the face from more than 40 million images, which can then be adapted to distinct health outcomes such as biological age and survival,” Aerts said.

THE TWO HEADS

• FAHR-FaceAge: Estimates biological age, meaning how old the face looks compared with the years on a birth record, after fine-tuning on ten public image sets, with a mean absolute error of 5.1 years.
• FAHR-FaceSurvival: Ranks death risk in people with cancer from the photograph itself, after fine-tuning on 34,363 Harvard radiation photos and testing on a held-out 3,820.
• Where they look: FaceAge leans on nasolabial folds (the lines from nose to mouth), the forehead above the eyes, and the temples. FaceSurvival leans on the area under the eyes and the bridge of the nose.

The two scores were only weakly correlated, and each still tracked survival when the other was in the model. Combining them improved the fit in an external Dutch cohort. That split is the design, not a bug. A face is being asked to yield more than one clinical number from the same pixels.

Patients Who Looked Older Had a 21% Higher Death Risk

Clinical testing used two radiation cohorts. Harvard-RT covers 38,183 patients treated at six Harvard-affiliated sites. Maastro-RT covers 4,892 patients treated in Maastricht, in the Netherlands. Survival models were adjusted for age, sex, race or ethnicity, cancer site, treatment intent, treatment year, and radiation technique, according to the published FAHR-Face validation study in The Lancet Digital Health.

The 21% figure is a group contrast over about four years of follow-up, not a personal chance of dying. A patient whose face looks five years older does not, on that basis, have a 21% probability of death.

HOW THE RISK SPLIT IN TESTING

Among the 38,183 Harvard patients, a third had a face-age gap of at least five years. Pratik Shah, a pathology professor at the University of California, Irvine, who directs its Machine Learning for Health Research Program and was not on the paper, said the facial-age model showed survival associations in 18 of 28 cancer categories and “may not perform consistently across cancer types.” The findings come from historical photos, he said, not from a trial that changed care.

The Survival Model Watches Under the Eyes

Raymond Mak, a radiation oncologist at Mass General Brigham Cancer Institute and corresponding author, said he was surprised the models could pick up mortality risk from a photograph at all. The age head, he said, looks at the temples, the forehead above the eyes, and the nasolabial fold. The survival head looks at a different zone: the nasal bridge and the area under the eyes.

Those are old landmarks. Mak noted that a Greek medical text from around 400 B.C. described sunken eyes and fallen temples as signs of wasting near death. The algorithms had no access to that literature. They were trained on photographs plus age or death labels.

To think that our algorithms might be replicating that wisdom of Hippocrates from millennia ago was mind-blowing to us.

Raymond Mak, radiation oncologist, Mass General Brigham Cancer Institute

Attention maps show where the network looks. They do not show what it measures. Mak said the maps cannot be translated into crow’s feet, eye bags, wrinkles, pigment, skin texture, or muscle wasting at the temples. “We don’t know exactly what these algorithms are looking at,” he said.

Lifestyle still left a mark on estimated age. Heavy alcohol use, drug use, and smoking were tied to looking biologically older, with the strongest gradient in smoking: current smokers older than former smokers, who looked older than people who had never smoked. Mak also said FaceAge held up across pose, lighting, expression, and makeup, and across most tested plastic surgery, with a small shift after rhinoplasty of about one year, which he linked to tissue loss rather than a cosmetic “tell.”

What the Algorithm Still Cannot Name

Aerts called FaceSurvival the first showing that an AI model can “directly predict cancer survival from a simple photograph.” That is a claim about a statistical ranking, not a reason. The paper does not identify a biological pathway, and it does not test whether acting on the score helps patients live longer or suffer less.

Miroslaw Bober, professor of video processing and machine learning at the University of Surrey, who was not involved, called the Dutch replication a real strength and the 90-day split in the US high-risk slice evidence that the tool can separate groups. He then drew a harder line.

Predicting an outcome is not the same as explaining it. Despite the references to ‘biological age’, the AI estimates age from facial appearance rather than directly measuring any underlying biological ageing process.

Miroslaw Bober, professor of video processing and machine learning, University of Surrey

WHERE EXPERTS DISAGREE

• On photo noise: Mak said pose, lighting, expression, and makeup had little effect. Bober said smiling or changing the lighting shifted the age estimate by around two years on average, and more in some cases, so claims about tracking ageing or lifestyle over time are premature.
• On race: The paper’s summary said both models generalised across age, sex, race, and cancer subgroups. Bober said the facial-age measure was not significantly associated with survival for Black patients after other factors were adjusted, though the separate survival model was.
• On clinic use: Mak’s stated goal is a quantitative, consistent, scalable assist, “rather than replace physician judgment.” Bober said the tool ranks relative risk inside a group and has not been validated as an individual survival probability or for decisions at the bedside.

Wilbur Lam, research chair and professor of pediatrics and biomedical engineering at Emory University, who was not involved, called the associations compelling and the clinical utility unproven until prospective tests cover specific facial features, skin color, and age groups. Bober added a gap the paper does not close: it did not test whether the network sees anything a doctor would miss, and it did not measure doctors’ own assessments of general health and fitness. “This is a promising proof of concept, but it is not ready for clinical use,” he said.

More Than 10,000 People Uploaded a Face for Research

The number is already leaving the radiation bunker. Mass General Brigham has opened the public FaceAge Research Study portal, an institutional review board-approved site where volunteers upload photos and receive a research readout of biological age. The team said more than 10,000 people have signed up since a quiet launch in March 2026.

Haugg, Mak, and Aerts are named on a US provisional patent, application 63/769,840, covering next-generation facial health algorithms. The European Research Council funded the paper. Aerts lists advising and consulting roles, unrelated to this work, for Onc.AI, Love Health, Sphera, Editas, AstraZeneca, and Bristol Myers Squibb. Mak lists advisory, consulting, honoraria, and research ties that include AstraZeneca, Varian, ViewRay, and Siemens.

FROM LAST YEAR’S MODEL TO A PUBLIC SCORE

1. May 8, 2025: FaceAge, trained on 58,851 presumed-healthy photos and tested on 6,196 cancer patients, is published as an age estimator linked to survival after radiation.
2. March 16, 2026: Haugg posts that the group is launching the IRB-approved public study so the biomarker can be tested in people who are not hospital patients.
3. October 19, 2026: FAHR-Face is published, adding a survival head trained on 34,363 cancer photos and tested on 3,820, with external check on 4,892 Dutch patients.
4. Now in clinic, not yet reported as results: A Brigham and Women’s trial enrolls 60 patients aged 65 and older with early-stage lung cancer before radiation, asking whether a photo can flag who needs extra nutrition or physical therapy. A second Mass General Brigham surgical trial pairs FaceAge with wearable-device data to flag complication risk.

Mak said most older patients with early-stage lung cancer have geriatric vulnerabilities, “but very few ever receive a formal geriatric assessment.” If a photograph can flag the most vulnerable, he said, that would be a practical way to close the gap. That is also how a research score becomes a triage object. The 60-person lung study is the first place the number may meet a real treatment plan, including supportive care some patients never get.

A survival rank that can be computed from a consumer photo will not stay inside radiation oncology by default. Insurers, employers, and consumer apps do not need a hospital badge to run a face through a model. The authors frame the portal as research, not diagnosis. The second effect is simpler: thousands of non-patients have already handed the lab a labeled face.

Last Year’s FaceAge Stopped at How Old You Look

The 2025 system, described on the lab’s page for the earlier FaceAge algorithm from 2025, estimated age from a photo and then used that age gap as a prognostic clue. Cancer patients in that study looked older than a non-cancer comparison group, and an older FaceAge tracked worse survival, especially when the estimate passed 85. Clinicians predicting short-term survival in a palliative radiation set improved when they were given the FaceAge output on top of a photo and chart details.

FAHR-Face keeps that age task and adds a second model that skips age and learns death labels from faces directly. Training scale is the other change: tens of millions of unlabeled faces first, then a few tens of thousands of labeled clinical photos. Aerts’s point is that a foundation model makes new facial biomarkers cheaper to train. The same structure means any bias in the shared representation can flow into every downstream score.

Mak’s clinic problem has not changed. Chronological age is a blunt instrument for deciding who can tolerate radiation, surgery, or drug treatment. A photograph is cheap. A full geriatric workup is not. The missing test is whether the new number adds anything once a physician has already looked at the patient, the labs, and a standard frailty scale. Until a prospective study measures that, the selfie is a ranking of past radiation patients, not a license to withhold care.

The 60-person lung trial will not settle fairness, consent, or what happens when a score leaks beyond the chart. It will show whether a photo can pull a few older patients into extra support. That is a narrower use than the headline, and it is the one the authors can actually run.

Frequently Asked Questions

What Photo Collection Was Used to Train FAHR-Face?

The foundation model was pretrained with self-supervised learning on unlabeled faces from WebFace42M. FaceAge was then fine-tuned in two age-balanced stages on ten public image datasets. FaceSurvival was fine-tuned only on the Harvard radiation photos, with a 90/10 split that kept 3,820 images out of training for the US test.

How Common Was a Five-Year Face-Age Gap in the Harvard Group?

Of 38,183 Harvard-RT patients, 12,610 (33.0%) had a face-age deviation of at least five years, and 6,506 (17.0%) had a gap of 10 years or more. The median deviation was 1.1 years, so the five-year threshold marks a large minority, not the typical patient.

Did the Researchers File a Patent on the Face Tools?

Mass General Brigham filed US Provisional Application No. 63/769,840 on next-generation facial health algorithms, with Haugg, Mak, and Aerts listed. A patent filing is not a regulator clearance, and the authors have not presented the system as a licensed medical device.

What Does the Public FaceAge Site Ask Volunteers to Do?

The FaceAge Research Study takes about 10 minutes. Participants sign a consent form, answer a survey that includes chronological age, willingness to share medical history, sleep, and exercise, then upload three to 10 selfies. The site returns a research estimate of biological age and lets people add later photos; the team says it is not a diagnosis.

Disclaimer: This article is news reporting on a published research study and is for information only. It is not medical advice, a prognosis, or a recommendation about cancer treatment, screening, or whether any person should upload a photograph to a research site. Readers who have questions about cancer care, biological age estimates, or trial enrollment should speak with their oncologist or another qualified physician who knows their history before acting on any score or headline figure. The numbers and study statuses here reflect the Lancet Digital Health paper, the Mass General Brigham release, and the named expert comments as of the dates on those documents, and they may change as trials report or as the models are revised.