Researchers Trained AI on 40 Million ‘Selfies’. What It Found About Cancer Survival Surprised Them

Radiation oncologist Dr. Raymond Mak and colleagues at Mass General Brigham built software that estimates a cancer patient’s odds of dying within 90 days from one picture of their face. Their results appeared in The Lancet Digital Health. Among 122 patients the model placed in its highest-risk group, 52.5% died inside that window. Among 975 patients in the lowest-risk group, roughly 2% did. The gap is wide enough that the authors argue facial appearance carries real information about a patient’s condition. Mak said he was surprised the signal existed at all, given that the input is a photograph and nothing else.
This article was created with the assistance of AI and reviewed by our editorial team for accuracy and clarity.

Cosmetic work did not throw the models off

An obvious objection is that faces get altered. Surgery changes bone-adjacent contours, injectables fill in creases, and makeup covers shadow and tone. Any of that could plausibly scramble a system reading faces for signs of physical decline. The research team checked, and reported that cosmetic surgery, injectables such as Botox, and makeup did not degrade the models’ performance. That result cuts in an uncomfortable direction. Whatever the software is picking up on appears to sit below the layer people spend money changing, in structure rather than surface finish.

Forty million photos came before the cancer data

The system started as a general-purpose model called FAHR-Face, short for Foundation Artificial Intelligence for Health Recognition-Face. It was trained on more than 40 million facial images covering a wide spread of ages and ethnic backgrounds. Only afterward did the team fine-tune it into two narrower tools. FAHR-FaceAge estimates biological age from one photo. FAHR-FaceSurvival predicts survival odds by combining facial features with other clinical information about the patient. Biological age is not the number of birthdays a person has had. It is an estimate of how far the body has worn at the cellular and physical level.

Twenty extra years of biological age, tracked

FAHR-FaceAge was tested by sorting patients according to the gap between their estimated biological age and their actual age. Among 1,060 patients the model scored as more than 20 years older than their chronological age, 21% died within 90 days. Among 4,195 patients scored at least 10 years younger than their chronological age, about 6% did. That is roughly a threefold difference in short-term mortality between the two ends. The sample sizes matter here: these are thousands of patients, not dozens, which is part of why outside readers called the statistical association robust even while disputing what it means.

Two algorithms, two different maps of the face

Mak described the models as reading separate territory. The age estimator concentrates on the area around the temples, the forehead, the region above the eyes, and the nasolabial fold, the line running from the nose down to the corner of the mouth. The survival model looks somewhere else entirely. According to Mak, it focuses on the bridge of the nose and the area beneath the eyes. The split was not designed in advance. Both models were trained on photographs paired with outcomes, so the regions they weight most heavily emerged from the data rather than from a clinician telling the software where to look.

Hippocrates described the same features around 400 B.C.

The facial regions the survival model weighted have a long documentary history. In the Book of Prognostics, written around 400 B.C., the Greek physician Hippocrates listed “the nose sharp, the eyes sunken, the temples fallen in” among the signs of muscle wasting and approaching death. That is close to the territory the algorithms settled on. Mak called the overlap eerie and said replicating that observation was mind-blowing to the team. He also stressed that no ancient text fed into the system. The models were trained purely on face photographs paired with age and with whether the person died, not on aggregated internet writing.

Prediction is not the same as explanation

Miroslaw Bober, professor of video processing and machine intelligence at the University of Surrey, drew a line through the study’s framing in a statement distributed by the Science Media Centre. He credited the work with convincing evidence of a statistical link between facial appearance and survival among cancer patients receiving radiotherapy. Then he objected to the vocabulary. Despite the references to biological age, he said, the system estimates age from facial appearance rather than measuring any underlying biological ageing process. A model that sorts patients accurately has still not told anyone why the sorting works, and the study did not test whether the software sees anything a physician would miss.

The comparison nobody ran was against doctors

The study’s central gap is what it did not measure. Clinicians already assess a patient’s general health and fitness before treatment, and those assessments were not recorded alongside the model’s scores. Without them, there is no way to show the software adds information a doctor did not already have. Bober called the work a promising proof of concept and said it is not ready for clinical use, specifically because it still needs testing against physicians’ own judgement and across different hospitals and patient groups. The authors agree on the broader point. They describe the models as an emerging tool, not a replacement for physicians or existing clinical evaluation.

Skin color and facial features remain open questions

Dr. Wilbur Lam, research chair and professor of pediatrics and biomedical engineering at Emory University, raised a separate concern by email. He called it an extremely compelling study drawn from a large data set, with associations that appear robust, then said its true clinical utility will need to be demonstrated prospectively. His sharper point concerned who the tool works for. Whether it can be accurately used for specific subpopulations remains a question, he said, naming people with particular facial features, skin color, and age groups as categories requiring prospective testing before anything gets implemented. Lam was not involved in the research.

Every patient studied was already a cancer patient

The scope limit is easy to miss. The survival findings come entirely from people already diagnosed and, in the radiotherapy cohort Bober referenced, already in treatment. Nothing in the study establishes what the models would do with a photograph of a healthy person. Mak said further research could reveal whether similar findings emerge among otherwise healthy people curious about their longevity, which is a hypothesis rather than a result. His stated near-term interest is narrower: catching accelerated aging or elevated mortality scores early enough to flag patients who might benefit from additional testing. For now, the tool identifies risk inside a population that already knows it is at risk.