AI estimates mortality risk in cancer patients from a facial photograph

Researchers in the United States have developed an artificial intelligence tool that analyzes a person’s face in a photograph to estimate biological age and predict short-term mortality risk for people with cancer. The model, named FAHR-Face, was trained on more than 40 million facial images and evaluated using data from over 30,000 cancer patients, the study reports.

How the tool performs

FAHR-Face combines two functions: FAHR-FaceAge estimates how old a person’s face appears compared with their calendar age, while FAHR-FaceSurvival generates a mortality-risk score. In the study, patients placed in the highest predicted-risk group had roughly a 50% chance of dying within 90 days, compared with about 2% in the lowest-risk group. The authors stress these figures describe group-level associations and do not translate into a definitive prognosis for an individual.

The algorithm inspects facial regions such as the area around the eyes, the forehead and the nose. The researchers say these zones may show signs that correlate with muscle wasting, changes in soft tissue or general deterioration in health, but they do not yet have a confirmed explanation for the exact visual cues the model uses.

"This is the first time we show that we can directly predict mortality risk — not just biological age, but also mortality risk — solely from a photograph," said Raymond Mak, a radiation oncologist at Mass General Brigham Cancer Institute and one of the study authors. He added that the aim is to make facial-photo–based health assessment "more quantitative, more consistent and scalable, to support personalized treatment decisions, rather than replace the clinician's assessment."

Independent experts urged caution. "That a photograph can predict survival does not mean it reveals something the clinician does not already know," warned Suchi Saria of Johns Hopkins University, noting that it remains to be shown whether adding facial analysis to existing clinical data actually improves prognostic accuracy. Miroslaw Bober of the University of Surrey highlighted a similar point: "Predicting an outcome is not the same as explaining it," and he warned the model is not validated as an individual-level survival probability and is not ready for clinical use.

There are additional concerns about fairness and generalisability. Because FAHR-Face was trained on large numbers of images, the model must be tested across diverse populations to ensure it performs equally accurately for people of different skin tones, ages and backgrounds. Researchers say further validation is needed before the tool could be considered as a supporting element in clinical decision-making rather than an experimental research method.

Photo: press material from the event