A selfie that predicts cancer survival: a promising frailty signal that is still far from a bedside tool

Researchers at Mass General Brigham have built an AI system that predicts a cancer patient's chances of survival from one ordinary photograph of their face. According to The Guardian, the results were published in Lancet Digital Health. The model behind it, FAHR-Face, was trained on more than 40 million facial photographs covering a range of ethnicities and ages.

The team built two tools on the same foundation model. FAHR-FaceAge estimates biological age, which the report describes as the age a person looks rather than their actual age, and uses that estimate to predict survival. FAHR-FaceSurvival skips the age step and predicts survival directly from the photo. The researchers say the two tools look at different parts of the face. FaceAge relies on the creases from the nose to the mouth, the forehead just above the eye sockets and the temples. FaceSurvival focuses on the area below the eyes and around the bridge of the nose. The headline figure comes from US patients treated with radiotherapy. Over an average follow-up of four years, those whom FaceAge judged to look five or more years older than their real age had a 21% higher risk of dying than the rest of the group. Those who looked five or more years younger had a 17% lower risk.

The technical idea matters more than any single result. Co-senior author Hugo Aerts describes FAHR-Face as a foundation model: a general model of the face learned once and then adapted to different health outcomes, so each new biomarker does not need to be trained from scratch. He calls it, to the team's knowledge, the first demonstration that AI can predict cancer survival directly from a simple photograph. The clinical need is real. Radiation oncologist Raymond Mak says that as populations age, doctors need better ways to judge physical reserve and frailty before recommending radiation, surgery or systemic therapy. Doctors already make that judgement by eye at every consultation. The study's ambition is to make it quantitative, consistent and scalable.

That is also where caution is needed. The reported numbers are associations at group level. They do not tell a specific patient how long they will live. A 21% difference in relative risk is useful for sorting patients by risk, but it is not a verdict on any one person. What the material shows is performance in particular study groups. Before this becomes a tool that shapes treatment, it needs prospective validation in other hospitals and populations. It also needs a clear check on whether accuracy holds equally across skin tones, ages and photo conditions, because a large training set does not automatically make a model fair. There is a second issue too: if a selfie is a health biomarker, then every face photo is potentially sensitive medical data. That concerns insurers and employers as much as hospitals.

Our reading: this is the kind of AI-in-medicine story that fits our long-term thesis. It turns cheap, everywhere-available data into clinical signal and can help personalise treatment and avoid treatments that do more harm than good. Combined with other biomarkers, it points toward medicine that is more preventive and better matched to each patient's real physical reserve, part of the path toward longer, healthier lives. In the short term, the right framing is a decision aid that supports the oncologist's judgement, not a replacement for it, and one that has to be regulated as health data from the start. The race now is not to find more things a face can predict. It is to show, in independent trials, that using these predictions actually improves patient outcomes.