Mass General Brigham develops the FAHR-Face model to predict survival risk of cancer patients using facial photos

This foundational AI model, called FAHR-Face, was developed by a research team at Mass General Brigham, a healthcare system in Massachusetts, USA, and was trained on more than 40 million facial photographs of people of different ages and ethnicities. Building on this, the researchers developed two complementary analysis tools: FAHR-FaceAge and FAHR-FaceSurvival. The former estimates a person's biological age and uses this information to help assess cancer prognosis; the latter directly predicts the survival risk of cancer patients from facial photographs, rather than first estimating age and then indirectly inferring the outcome.

The research team trained the models using clinical photographs of cancer patients and their subsequent medical outcomes. FAHR-FaceSurvival, in particular, was trained on 34,363 facial photographs of cancer patients and validated in independent patient datasets. The researchers hope in this way to enable artificial intelligence to identify signals in facial images that are related to patients' physical condition and survival outcomes.

The results show that FAHR-FaceSurvival has strong risk-discrimination ability. According to the model's predictions, about half of the patients in the highest 10% risk-of-death bracket died within 90 days, whereas this proportion was about 2% among patients in the lowest 10% risk bracket. When compared by quartiles, the mortality risk of the highest-risk group was also significantly higher than that of the lowest-risk group, with an adjusted hazard ratio of 3.23.

The researchers also found that the gap between the biological age estimated by artificial intelligence and actual age is likewise associated with the survival outcomes of cancer patients. Among patients receiving radiotherapy, those who looked at least 5 years older than their actual age had a 21% higher risk of death than other patients during a mean follow-up of about 4 years, while those who looked at least 5 years younger than their actual age had a 17% lower risk of death.

This does not mean that a person who looks young is necessarily better able to resist cancer. The biological age in the study is an indicator estimated by artificial intelligence based on appearance, not a direct measurement of cellular aging, organ function, or immune system status. It may comprehensively reflect bodily wear, changes in muscle and soft tissue, and other health-related factors, but exactly which specific physiological signals the model identifies remains incompletely understood.

By analyzing the model's attention regions, the researchers found that the two tools focus on different facial locations. FAHR-FaceAge mainly analyzes the nasolabial folds from the sides of the nose to the corners of the mouth, the forehead, and the temple regions; FAHR-FaceSurvival focuses more on the area below the eyes and around the bridge of the nose. This suggests that predicting biological age and predicting survival risk may rely on different facial information.

The research team believes that in the future this technology may help doctors better assess patients' physiological reserve and frailty before deciding whether to use surgery, radiotherapy, or systemic drug therapy. For older patients with complex medical conditions, or those lacking complete traditional biomarker data, facial photograph analysis could become a low-cost, easy-to-operate auxiliary assessment method.

The study also found that lifestyle is associated with the biological age presented by the face, with smoking showing the most prominent association with looking older, and excessive alcohol consumption and drug use also being associated with similar phenomena. The researchers also stated that FAHR-FaceAge shows a certain stability under different shooting angles, lighting, expressions, and makeup conditions, and maintained relatively stable estimation results after most of the plastic surgeries tested, although rhinoplasty caused a small change.

Nevertheless, the researchers and external experts emphasize that this technology cannot yet directly replace doctors' judgment. The accuracy of model predictions does not automatically mean that it can improve actual treatment outcomes. Future research still needs to confirm whether adding facial analysis results to the medical history, examination data, and physical fitness assessments already available to doctors can further improve predictive accuracy and truly help patients achieve better treatment outcomes.

In addition, the patient population used for model training mainly comes from high-income countries, and how to ensure its applicability to different ethnic groups, age groups, and healthcare settings remains an important issue. Because facial features may be related to racial background and prior medical conditions, artificial intelligence may also inherit biases in the training data. Therefore, broader independent validation and prospective clinical studies are still needed before large-scale application.

The related research findings have been published in The Lancet Digital Health. The research team hopes that in the future more quantifiable health indicators can be extracted from an ordinary photograph, making facial analysis a supplementary tool for clinical risk assessment, rather than deciding patients' treatment opportunities based on appearance alone.