Trained on 40 million faces: FAHR-Face estimates age and survival risk from a single photo
Researchers at Mass General Brigham have built a foundation model trained on more than 40 million facial images and adapted it into two tools: one that estimates biological age and one that predicts cancer survival directly from a single photograph. The study is published in Lancet Digital Health – but the tools are not in clinical use, and the researchers stress that the goal is to support, not replace, the physician's assessment.
Researchers at Mass General Brigham, a non-profit academic health system in Massachusetts, have developed an AI system that can estimate biological age and predict survival in cancer patients from a single facial photograph. The study behind the tools is published in the journal Lancet Digital Health and was covered by The Guardian, The Mirror and in a press release from the institution on 19–20 October 2026.
From foundation model to two tools
The system is called FAHR-Face, short for Foundation Artificial Intelligence for Health Recognition. It is a so-called foundation model trained on data from more than 40 million facial photographs spanning a wide range of ethnicities and ages, according to the press release from Mass General Brigham (Medical Xpress).
The foundation model was then adapted into two specialised tools. FAHR-FaceAge estimates a person's biological age from the face, while FAHR-FaceSurvival predicts survival risk in cancer patients directly from the photograph.
According to The Guardian (theguardian.com), this is "one of the most extensive demonstrations to date" of how AI analysis of a selfie can help assess a patient's health and potentially guide medical treatment.
The numbers behind the study
The FaceAge tool – which The Mirror describes as the precursor to the tools in the study – was trained on data from 58,851 presumed healthy people aged 60 or older. The researchers then examined data from 6,196 patients with cancer diagnoses from two institutions in the Netherlands and the USA, The Mirror reports (mirror.co.uk).
The results reported from the study include, among others:
- Among cancer patients treated with radiotherapy in the USA, those who according to FAHR-FaceAge looked five or more years older than their actual age had a 21 percent higher risk of dying during an average follow-up period of four years than the rest of the group. Those who looked five or more years younger had a 17 percent lower risk.
- Cancer patients looked on average about five years older biologically than their chronological age, compared with people without the disease.
- Smokers had a facial biological age about three years higher than non-smokers. According to the study's authors, there was no significant difference between former smokers and those who had never smoked.
The researchers also found that the two tools' predictions were only weakly correlated, but that both remained independently associated with survival when analysed together. Combining the two measurements additionally improved the models' fit in an external validation cohort of cancer patients, Mass General Brigham reports (Medical Xpress).
The two tools read the face differently
According to The Guardian, the two algorithms focus on different parts of the face:
- FAHR-FaceAge places weight on the nasolabial folds – the folds running from the sides of the nose down to the corners of the mouth – the forehead just above the eye sockets, and the temples.
- FAHR-FaceSurvival assesses the area under the eyes and around the bridge of the nose.
That the tools read different regions and yield weakly correlated yet individually significant predictions is part of what makes the combination meaningful in the study's data.
The researchers: support, not replacement
Hugo Aerts, one of the study's two senior authors, highlights in the press release that this is something new. "FAHR-FaceSurvival advances the field of AI-based facial healthcare by learning survival risk directly from facial photographs of cancer patients," he says, adding: "As far as we know, this is the first demonstration that an AI model can predict cancer survival directly from a single photograph" (Medical Xpress).
Raymond Mak, corresponding author and radiation oncologist at Mass General Brigham, stresses that the tool is intended as an aid for physicians. "Our goal is to make health assessment based on facial photographs more quantitative, consistent and scalable, so that it can support individualised treatment decisions rather than replace the physician's assessment," he says.
Not a clinical tool – yet
It is worth holding on to what this actually is: descriptions of the researchers' own results, conveyed through press releases and press coverage of the study. The tools are not in clinical use, and the sources point to a potential to contribute in treatment assessments – not to an established implementation.
Several caveats apply. The figures are presented by the study's authors via secondary sources. The tools predict risk at the group level in the study's cohorts – they cannot "predict" individual outcomes with certainty, and the researchers themselves emphasise that the goal is to supplement clinical judgement.
Open questions remain: how the tools perform on patient groups other than radiotherapy-treated cancer patients, what mechanisms underlie the association between facial biology and survival, and how robust the predictions are across populations. It is also unclear in the coverage how The Mirror's validation figures (58,851 healthy and 6,196 cancer patients) relate to the 40 million images the foundation model was trained on – likely these refer to different steps in the development, but the sources do not give a full picture.
The study is nevertheless a concrete example of how large general image foundation models can be fine-tuned to specific clinical questions – here biological age and survival – and of how the research itself deliberately keeps its distance from clinical use. The question going forward is what will be required before such tools can safely be taken into use in hospitals.