A team of American researchers has developed an AI capable of estimating a person’s biological age from their face and, in cancer patients, assessing their short-term survival chances. This work, published on October 19, 2026, in The Lancet Digital Health, comes from Mass General Brigham, the largest hospital-university network in Massachusetts, affiliated with Harvard.
In a cancer patient, the face can provide oncologists with valuable clues. These morphological markers help clinicians judge, for example, whether a patient is strong enough to withstand heavy chemotherapy or radiotherapy. However, this assessment is inherently subjective: it can vary from one practitioner to another and is not reproducible.
The Mass General Brigham team developed a model called FAHR-Face, standing for “Foundation Artificial Intelligence for Health Recognition – Face.” It was trained on over 40 million facial photographs, then specialized in two tasks: estimating biological age (FAHR-FaceAge) and predicting survival of a cancer patient (FAHR-FaceSurvival). According to the authors, this is “the first demonstration that an AI model can directly predict the survival of a cancer patient from a simple photograph.”
A “foundation model” focused on medicine
Like ChatGPT or Gemini, FAHR-Face is a “foundation model”, meaning a model first trained on a very large volume of data to acquire a general understanding of its domain, then specialized for targeted tasks. The difference lies in its corpus: photographs of human faces, of all ethnic origins and ages, rather than text. By analyzing them, the model learned what structures a face (proportions, skin texture, depth of folds, condition of the suborbital tissues), but also what alters it over time, such as cellular aging, pathologies, smoking, alcohol, and nutritional deficiencies. The researchers then directed this foundation toward two medical applications.
FAHR-FaceAge focuses on the actual condition of a person’s cells, tissues, organs, and physiological systems, as opposed to their chronological age. It concentrates on the forehead, temples, nasolabial folds (the creases connecting the nose to the corners of the lips), and the eye contours. Validated on a cohort of 38,000 Harvard patients, it indicates that those whose faces appeared at least five years older than their chronological age had a 21% higher risk of death, from all causes combined, after adjusting for age, sex, cancer type, and disease stage. The model also highlighted an association between smoking, alcohol, drug use, and accelerated facial aging.
FAHR-FaceSurvival assesses the risk of death for a cancer patient by analyzing the suborbital area and the nasal bridge perimeter. To test it, researchers submitted photographs of patients treated with radiotherapy to the model, providing no other information than their face. The model classified each patient into a risk category, from lowest to highest.
Among the 122 patients placed in the most critical category, 52.5% actually died within 90 days following the photograph. In the lowest-risk category, which included 975 patients, about 2% died within the same period. By dividing patients into four groups according to prognosis severity, those in the highest-risk group died 3.23 times more often than those in the least exposed group.
A retrospective study, without real-world trials
These results must be put into context. The study is retrospective: the model was fed photos of patients whose outcomes were already known, then its predictions were compared afterward to reality. It shows that FAHR-FaceSurvival correctly classifies past data, but it does not yet demonstrate that it would improve medical decision-making if used in real time. Only prospective clinical trials, integrating the tool into the care pathway, would measure its concrete effect. Two such trials have already been launched at Mass General Brigham.
The researchers also opened a public portal where anyone can upload a photo of their face to enrich their database. According to the homepage, the procedure takes no more than 10 minutes, and “all photos and data are stored securely and used solely for research purposes.” In return, the user receives an estimate of their biological age by FAHR-FaceAge, and their participation remains anonymous.
