Mass General Brigham's face photo AI triples mortality risk spread in cancer patients

Researchers led by Dr Raymond Mak at Mass General Brigham in Boston have published a foundation model that reads a single face photograph to estimate biological age and predict survival in people with cancer. The model development and validation study, FAHR-Face, appeared in the Lancet Digital Health and was funded by the EU's European Research Council ¹².

What the study found

The team trained the base model on more than 40 million facial photographs, then tuned it for two tasks. FAHR-FaceAge, which estimates biological age, was fine-tuned on ten public image datasets. FAHR-FaceSurvival was fine-tuned on 34 363 photographs of patients with cancer ². Most of those patients came from the US and the Netherlands and were in their 50s or 60s, although the range ran from children to people in their 90s ¹.

On public datasets FAHR-FaceAge had a mean absolute error of 5·1 years, the lowest of the models it was compared with, and it held its accuracy across the whole lifespan ². FAHR-FaceSurvival separated patients clearly by risk: those in the highest-risk quartile had more than three times the mortality of the lowest, with an adjusted hazard ratio of 3·23 (p<0·0001). The authors adjusted survival analyses for clinical prognostic factors and tested both models in two independent cancer cohorts, where the results held across age, sex, race and cancer subgroups ². The two models gave complementary information, and using them together improved prognostic accuracy ².

Attention mapping showed the survival model leaned most on the area under the eyes and around the bridge of the nose, across cancers of varying severity, including disease that had spread to bone or brain ¹. The team also tested how well the models held up against make-up, pose, lighting and cosmetic surgery. Mak said the age tool was “surprisingly consistent” across real-world photos, with “only a small shift observed after rhinoplasty” ¹.

What it means for the market

The authors argue that one foundation model can produce cheap, scalable facial biomarkers, and that it allowed effective training on relatively small clinical datasets ². That second point matters to developers of clinical AI, for whom labelled patient data is usually the scarce input. The stated use is risk stratification where conventional biomarkers are not available ². Mak framed the aim as helping clinicians judge physical reserve and frailty before intensive treatment such as radiation, surgery or systemic therapy, not replacing their judgement ¹.

The tool is not a product yet. A patent application is pending, and the team is collecting photos from the public in return for a personal biological age estimate, a route to the larger and more diverse data it says it needs ¹. For NHS and European buyers, the relevant questions are evidence and regulation rather than procurement: the study gives no clinical deployment, no regulatory submission and no comparison with clinicians.

Caveats

The BBC describes the findings as preliminary and the tools as still under development and far from ready for hospital use ¹. The researchers say they need many more photos, particularly from a wider range of ethnic groups, to judge how well the tools perform ¹. Prof Miroslaw Bober of the University of Surrey, who was not part of the team, called it “a promising proof of concept, but it is not ready for clinical use” and said it still needs testing against doctors' own judgement and across different hospitals and patient groups. He added: “Predicting an outcome is not the same as explaining it.” ¹