AI facial tools estimate cancer survival, but remain unproven in care

Two tools, one facial model

FAHR-FaceAge estimates biological age — how old a person looks — and uses it to predict survival. FAHR-FaceSurvival predicts survival directly from a photograph, without biological age as an intermediate measure. Both use the same underlying model, FAHR-Face, trained on more than 40 million images.

The tools also attend to different facial regions. FAHR-FaceAge focuses on the folds running from the nose to the corners of the mouth, the forehead above the eye sockets, and the temples. FAHR-FaceSurvival looks at the area beneath the eyes and around the bridge of the nose.

Hugo Aerts, one of the study’s senior authors, said the direct-survival tool extends facial health recognition by learning to estimate the risk of death from photographs of cancer patients. The researchers call their work the first example known to them of an AI model directly predicting cancer survival from an ordinary photograph.

Aerts said the shared model can be adapted to different health measures, including biological age and survival. That could let researchers build several facial biomarkers without training each model from scratch.

A signal, not a treatment decision

The age estimates were associated with outcomes in the study: among US patients receiving radiation therapy, looking five or more years older corresponded to a 21% higher risk of death over a median four-year follow-up, while looking five or more years younger corresponded to a 17% lower risk. The results also suggest biological-age estimates are closely tied to physical health, potentially reflecting a patient’s ability to tolerate treatment.

Radiation oncologist Raymond Mak of Mass General Brigham said clinicians need better ways to assess patients’ physical reserve and frailty before radiation, surgery, or systemic therapy. The aim, he said, is to make photo-based health assessment quantitative, consistent, and scalable — while keeping treatment decisions with the physician.

The researchers also found that heavy alcohol use, drug use, and smoking were associated with older AI-estimated biological age. Mak said FAHR-FaceAge produced stable results across ordinary differences in head position, lighting, facial expression, and makeup. Estimates also remained stable after most of the plastic surgeries tested, with a small change only after rhinoplasty.

What a selfie cannot settle

I think the most important question is not whether a face contains a health signal, but what clinicians should do with it. The study points toward a low-cost way to support personalized treatment choices, yet an association with survival does not establish that the score can safely guide an individual patient’s care.

The announcement is quiet about how much the tools improve decisions over existing assessments, or where a useful threshold would fall. Until further research answers that, a facial score may help prompt a closer look at a patient’s condition, but it cannot carry the weight of a treatment recommendation.