Researchers at Mass General Brigham have introduced two artificial intelligence tools that can estimate a person’s biological age and predict mortality risk in cancer patients using nothing more than a facial photograph. The models, named FAHR-FaceAge and FAHR-FaceSurvival, are detailed in a study published in The Lancet Digital Health. The work builds on FAHR-Face, a foundation model trained on more than 40 million facial images from the WebFace42M dataset, which learned visual patterns by reconstructing partially obscured photos—a technique known as masked autoencoding.
- Foundation model: FAHR-Face was pre-trained on over 40 million faces, then fine-tuned into two specialized tools.
- Biological age estimation: FAHR-FaceAge, refined on 10 public databases, estimates biological age with an error of roughly five years.
- Survival prediction: FAHR-FaceSurvival was trained on 34,363 photographs of cancer patients and directly predicts survival outcomes.
- Clinical validation: Both models were tested on independent cohorts, including nearly 5,000 radiotherapy patients in the Netherlands.
How the Models Read the Face
Each tool focuses on distinct facial regions. FAHR-FaceAge examines the nasolabial folds, the forehead just above the orbits, and the temples. FAHR-FaceSurvival, by contrast, analyzes the area beneath the eyes and around the bridge of the nose. According to Dr. Raymond Mak, a radiation oncologist and corresponding author, the survival algorithm “looks at the bridge of the nose and the area underneath the eyes” to draw conclusions about a patient’s prognosis.
The hospital emphasizes that the age estimator is not easily fooled by posture, lighting, facial expression, or makeup, nor by most cosmetic surgeries—though it does deviate slightly after a rhinoplasty. Smoking emerged as the habit most strongly associated with appearing biologically older, ahead of heavy alcohol consumption and drug use.
What the Study Found
The results are striking at the group level. Patients placed in the highest-risk quartile by FAHR-FaceSurvival had significantly higher mortality than those in the lowest-risk quartile, with an adjusted hazard ratio of 3.23. In one analysis, 52.5% of 122 patients classified in the highest-risk group died within 90 days, compared with about 2% of 975 patients in the lowest-risk group. The gap between estimated biological age and chronological age was also independently associated with survival: patients who looked five or more years older than their actual age had a 21% higher risk of dying during an average four-year follow-up, while those who looked five years younger had a 17% lower risk.
The two tools barely correlate with each other, and combining them improved the fit of the survival model in an external cohort. Hugo Aerts, director of the hospital’s AI in Medicine program and senior co-author, framed the work as the first demonstration that an AI model can predict cancer survival directly from a photo. Mak stressed that the goal is to make clinical assessment more quantitative and consistent—supporting, not replacing, physician judgment.
“There was actually an internal debate in our lab, should we even do this research? The potential for misuse is so great,” Mak said. His team ultimately decided to develop the technology in the public eye so it could be overseen by regulators and ethicists.
Not Ready for the Clinic
The study is a proof of concept, not a finished clinical product. Miroslaw Bober, an image and AI specialist at the University of Surrey, noted that the research did not test whether the AI detects something a physician would miss, nor whether its predictions add useful information beyond clinical assessments of health and fitness—which were not measured. He described it as a promising concept that is not yet ready for clinical use. Dr. Wilbur Lam of Emory University added that performance needs further evaluation across patient subgroups, and prospective studies are required to show whether the tool actually improves care.
Two other details merit attention. Mass General Brigham has filed a provisional patent application in the U.S. covering these facial health algorithms. The team has also launched a public study portal where anyone can upload a photo, receive an estimate, and contribute data to validate the model.
The technology follows a broader trend of clinical AI that must move from strong retrospective performance to real-world decision-making. Measuring is not the same as diagnosing, and a model that turns a selfie into a mortality risk has obvious applications outside the hospital—starting with insurers. Before celebrating that a photo is enough, two demands are essential: demonstrate that the tool improves clinical decisions, and establish clearly who may use it and with what data.