Insights

Can a selfie predict your cancer survival?

Close-up of a face with a grid of light and shadow across the eyes and nose

A new study in The Lancet Digital Health describes an AI system that reads a single face photograph and produces two separate health readings: an estimate of biological age and a direct prediction of survival in patients with cancer. It comes from the Boston group (Mass General Brigham) behind FaceAge, the tool reported last year.

How it was built

The system, called FAHR-Face (Foundation Artificial Intelligence for Health Recognition), is a foundation model: a masked autoencoder pre-trained with self-supervised learning on more than 40 million unlabelled face images. At that stage the model knows nothing about health. It simply learns what faces look like across the whole lifespan.

The team then fine-tuned it for two tasks. The first, FAHR-FaceAge, was retrained on ten public image datasets to estimate biological age. Biological age is a measure of how fast your body is functioning at a cellular and systemic level, rather than the number of years since you were born. On a public benchmark of presumably healthy people it had an average error of 5.1 years, the lowest of the models compared, and it stayed accurate from childhood to old age.

The second, FAHR-FaceSurvival, was trained on 34,363 photographs of cancer patients, mostly routine pictures taken before radiotherapy, linked to what later happened to them. It skips age altogether and learns to estimate death risk directly from the face.

What they found

In cancer patients, a larger gap between the age the AI estimated and the person’s real age was linked to poorer survival, even after adjusting for age, sex and cancer type, and the new tool beat the team’s earlier FaceAge model.

The survival tool’s results were more striking. Patients in its highest-risk quarter had an adjusted hazard ratio for death of 3.23 compared with the lowest-risk quarter. The difference was sharpest in the short term, with roughly half of those flagged as highest risk dying within 90 days, against about 2 in 100 of the lowest-risk group. The findings were re-checked in an independent cohort of 4,892 patients from the Netherlands and held across age, sex, race and cancer subgroups.

Which parts of the face matter

The researchers used attention maps to see which parts of the face each model relied on, and the two tools looked at different places. The survival model focused on the area just below the eyes and the bridge of the nose. The age model focused on the smile lines (nasolabial folds), the forehead above the eye sockets and the temples.

The two scores were only weakly correlated (R = 0.16). Each remained independently associated with survival when analysed together, and combining them improved prediction.

The survival model’s focus is intriguing to anyone who has trained in clinical medicine. It closely matches Hippocrates’ description, from around 400 BC, of the face of a dying patient: a sharp nose, hollow eyes and sunken temples. Senior author Raymond Mak suspects the model may be detecting loss of soft tissue and muscle, though he stresses this is speculation. The AI was given only photographs and outcomes, with no medical texts or clinical notes.

Lifestyle, robustness and make-up

Heavy drinking, drug use and smoking were associated with looking biologically older. Smoking showed a clear gradient: current smokers looked older than former smokers, who looked older than people who had never smoked. This is not a surprising finding, as we know these factors affect biological ageing and longevity.

A common worry is that these tools can be fooled. The age model stayed consistent across pose, lighting, expression and make-up, and across most plastic surgery procedures tested. The one noticeable shift came after rhinoplasty.

My view

I’m not surprised that a photograph carries prognostic information. Oncologists and other physicians make these judgements every day, based on our clinical experience. Performance status, clinical frailty scores, exercise tolerance, grip strength and a patient’s general appearance in clinic all feed into whether, and how intensively, we treat them. In that sense, “looking unwell” has always been part of our assessment, and this tool tries to measure it objectively.

The limitations matter. The analyses adjusted for age, sex, race, cancer type and treatment, but not for detailed frailty measures such as grip strength, walking speed or weight loss. The model may therefore be picking up frailty that we can already assess at the bedside. This was historical data rather than a clinical trial. A face also reveals ethnicity, so the model could learn patterns that reflect unequal diagnosis or care rather than biology, and the patients came mainly from high-income countries. The authors call for testing in broader populations and prospective trials before any clinical use.

I would never rely on AI alone for a decision of this weight. Treatment choices depend on the patient’s goals, their other health conditions and the biology of the tumour, none of which are evident in a photograph. The authors of this study hope this AI tool can be used to support clinical decision making in the future. The question is whether a facial signal adds anything to the tools we already use, and whether acting on it improves outcomes for patients. Until that has been tested properly, this is a fascinating research tool, not a clinical one.

Read the full study: Foundation artificial intelligence models for health recognition by use of face photographs (FAHR-Face), The Lancet Digital Health, 2026.