An AI that reads cancer survival odds from a selfie is promising, but its value is as a frailty gauge, not a verdict

Researchers at Mass General Brigham have built an AI system that predicts a cancer patient's chances of survival from a single ordinary photograph of their face. The work, published in Lancet Digital Health, uses a foundation model: a general-purpose model trained once on a large dataset and then adapted to specific tasks. This one was trained on more than 40 million facial photographs across a range of ethnicities and ages. According to The Guardian's report, it is one of the most comprehensive demonstrations yet of selfie-based health assessment.

The team built two tools from the same model. FAHR-FaceAge estimates biological age, meaning how old a person looks rather than their age in years, and uses that to predict survival. FAHR-FaceSurvival predicts survival directly. The two rely on different parts of the face. FaceAge looks at the creases from nose to 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 result comes from US cancer patients treated with radiotherapy and followed for four years on average. Those whom FaceAge judged to look five or more years older than their actual 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. Co-senior author Hugo Aerts calls it, to his knowledge, the first demonstration that AI can directly predict cancer survival from a simple photograph.

The clinical purpose is specific. Radiation oncologist Raymond Mak says doctors need better ways to estimate 'physical reserve and frailty' before recommending intensive treatments such as radiation, surgery or systemic therapy. Today that judgment is largely the 'eyeball test,' an experienced clinician's impression of how robust a patient looks. The goal is to make that impression 'quantitative, consistent and scalable.' That is a modest and sensible ambition: putting a number on something doctors already do informally.

The caveats are real. A 21% difference in relative risk is a statistical signal across a population, not a prediction for one person. The reported results come from one patient group, radiotherapy patients in the US, so performance elsewhere needs independent checking. Faces also carry ethnicity, lighting and cosmetic variation, which raises fairness and robustness questions. The diverse training data addresses these only partly. Then there is privacy. A model that infers mortality risk from a face could be valuable to insurers or employers. It needs firm limits on use outside clinical care, with patient consent.

Our reading: this fits a pattern we keep seeing in medical AI. The value comes from extracting new signal from data that already exists or costs almost nothing to collect, here a photo, with no new test required. Aerts's foundation-model point is the strategic part: one learned representation of the face can produce many health biomarkers without training each from scratch. That is how AI will steadily extend healthy life, through many cheap, scalable measurements that help doctors treat the right patient with the right intensity. The tool should inform the clinician's decision, not make it. It also needs to clear validation and governance hurdles before it reaches a clinic.