AI Model Estimates Cancer Survival From Facial Photographs, Separating High- and Low-Risk Patients

BOSTON — Researchers at Mass General Brigham have developed an artificial intelligence system that can estimate a cancer patient's risk of dying from a single facial photograph, according to a study published this week in The Lancet Digital Health.

The system, called FAHR-Face (Foundation Artificial Intelligence for Health Recognition), was trained on more than 40 million facial images spanning a wide range of ages and ethnicities. From that foundation model, the team created two tools. FAHR-FaceAge estimates biological age — a measure of how old a person appears physiologically rather than their chronological age — with a mean absolute error of about five years, which the researchers said is more accurate than previous age-estimation algorithms. FAHR-FaceSurvival predicts survival outcomes directly from the photograph.

In testing on cancer patients, the survival model showed a clear separation between risk groups. Among 122 patients assigned the highest risk scores, 64 (52.5 percent) died within 90 days of the photograph being taken. In the lowest-risk group of 975 patients, only about 2 percent died in the same period. Over longer follow-up, patients in the highest-risk quartile had more than three times the mortality risk of those in the lowest quartile, even after adjustment for factors including age, sex, cancer type and treatment characteristics. The association held in independent validation cohorts.

The biological-age tool produced a related signal. Cancer patients whose faces appeared at least five years older than their chronological age had a 21 percent higher risk of death than others. Those who appeared at least five years younger had a 17 percent lower risk. Among patients whose estimated biological age was more than 20 years above their chronological age, 21 percent died within 90 days, compared with roughly 6 percent of those who looked at least 10 years younger.

Hugo Aerts, director of the Artificial Intelligence in Medicine program at Mass General Brigham and a co-senior author of the study, said the work advances facial analysis for health applications. "FAHR-FaceSurvival advances the field of facial health recognition AI a step further by learning survival risk directly from facial photographs of patients with cancer," Aerts said. "To our knowledge, this is the first demonstration that an AI model can directly predict cancer survival from a simple photograph."

The two tools focus on different facial regions. FAHR-FaceAge emphasizes nasolabial folds, the forehead area above the eyes and the temples. FAHR-FaceSurvival evaluates the area below the eyes and around the nasal bridge. Their predictions were only weakly correlated, and combining them improved model performance.

Raymond Mak, a radiation oncologist at Mass General Brigham Cancer Institute and a study author, noted that physicians already assess patients visually but often struggle with precise prognosis. "Doctors are trained to assess patients just by how they look, but they are in general not very good at estimating end-of-life expectancy," Mak said. "We think that using AI, we can capture some of that physiological reserve, risk for mortality, just from a face photo."

The survival model was trained on photographs of roughly 34,000 cancer patients treated with radiotherapy, along with their outcome data, and tested on held-out groups totaling several thousand additional patients in the United States and the Netherlands. Lifestyle factors such as heavy alcohol use, drug use and smoking were associated with older-appearing faces in related analyses.

Researchers described the findings as preliminary. The tools are not yet ready for clinical decision-making, and further prospective validation across diverse hospitals and patient populations is needed. A publicly accessible research website allows individuals to test the age-estimation component with a selfie as part of an ethics-approved project, but experts caution that the technology is experimental.

The study authors said facial analysis could eventually serve as a low-cost, non-invasive supplement to existing clinical information when physicians evaluate treatment intensity or supportive care needs. No such routine use has been established.

The research appears in the journal The Lancet Digital Health.