An AI that reads cancer survival from a face photo

The system is called FAHR-Face (Foundation Artificial Intelligence for Health Recognition). It starts as a general-purpose face model trained on more than 40 million face photographs from the public WebFace42M dataset.

From that base, the team built two tools. FAHR-FaceAge estimates a patient's biological age — how old the body looks versus the birth certificate — with a reported error margin of around five years, better than prior age-estimation algorithms. FAHR-FaceSurvival goes further: trained on 34,000 facial images of people with cancer, it predicts survival risk directly from the photo, without going through biological age as an intermediate step.

Crucially, the two tools are not the same signal twice. Their predictions were only weakly correlated, they read different parts of the face (FaceAge keys on the nose-to-mouth creases, forehead, and temples; FaceSurvival looks below the eyes and around the bridge of the nose), and combining both improved the survival model's fit in an external validation cohort.

The paper's headline number: the highest-risk quartile had more than triple the mortality of the lowest, an adjusted hazard ratio of 3.23. The raw 90-day counts make it visceral: in the held-out test set of 3,820 patients, 64 of the 122 patients (52.5%) in the top score band (0.9-1.0) died within 90 days, versus 20 of 975 (2.1%) in the bottom band (0.0-0.1). The findings were validated in an independent cohort and held across age, sex, race and cancer subgroups.

The biological-age angle told a consistent story on its own. In radiotherapy patients, those who looked five or more years older than their actual age carried a 21% higher risk of dying over a median follow-up of 4.3 years; those who looked five or more years younger carried 17% less risk. The age gap was independently linked to survival even after accounting for age, sex, race, and cancer type.

The validation wasn't one hospital's patients: the main cohort drew on 38,183 patients treated with radiotherapy at six Harvard-affiliated sites, and the external validation set covered 4,892 cancer patients treated with radiotherapy in Maastricht, the Netherlands. The researchers report FAHR-FaceAge held up across common real-world photo variations — pose, lighting, expression, and makeup. Only rhinoplasty among the tested plastic surgeries produced a small shift.

Doctors already size up how frail a patient looks before recommending punishing treatments — the AI tries to make that judgment measurable and repeatable. As corresponding author Raymond Mak, a radiation oncologist at Mass General Brigham, put it: the goal is "to make face photograph-based health assessment more quantitative, consistent and scalable so it can support personalized treatment decisions rather than replace physician judgment."

That framing matters. This is a research system, not a clinical product — no FDA clearance, no deployment in clinics. Mass General Brigham has filed a provisional patent on the next-generation facial algorithms and is expanding the work through a public, IRB-approved online study at faceage.bwh.harvard.edu, where participants can upload a photo and receive their own FaceAge research readout.

Outside experts urge caution: Johns Hopkins' Suchi Saria told the Washington Post it isn't yet known what physiological signal the model reads, and UC Berkeley's Emma Pierson warned that face-trained models implicitly include race and could encode racial bias.

The model's lifestyle findings pass the smell test: heavy alcohol use, drug use, and smoking were all associated with looking biologically older, with smoking showing the strongest pattern — current smokers looked older than former smokers, who looked older than never-smokers.

"To our knowledge, this is the first demonstration that an AI model can directly predict cancer survival from a simple photograph," said co-senior author Hugo Aerts, who directs Mass General Brigham's Artificial Intelligence in Medicine program. The full paper is "Foundation artificial intelligence models for health recognition by use of face photographs (FAHR-Face): a model development and validation study," led by first author Fridolin Haugg.