Published Oct 22, 2026 | 7:00 AM ⚊ Updated Oct 22, 2026 | 7:00 AM
Researchers said that image quality, lighting and facial expression could affect accuracy. (Representational image/iStock)
Synopsis: An AI model trained on more than 40 million facial photographs has identified patterns associated with biological ageing and mortality risk among cancer patients. Published in ‘The Lancet Digital Health’, the study suggests facial analysis could complement clinical assessments, but researchers say broader validation and prospective trials are needed before routine use.
An artificial intelligence model trained on more than 40 million facial photographs has identified patterns associated with biological ageing and mortality risk among cancer patients, suggesting that a simple photograph could one day complement doctors’ assessments before intensive treatment.
Developed by researchers at Mass General Brigham, the model, called Foundation Artificial Intelligence for Health Recognition-Face (FAHR-Face), powers two tools: FAHR-FaceAge, which estimates age from facial features, and FAHR-FaceSurvival, which ranks cancer patients by their predicted survival risk.
In a study published in The Lancet Digital Health, patients placed in the highest-risk quartile by FAHR-FaceSurvival had more than three times the hazard of death compared with those in the lowest-risk quartile. The adjusted hazard ratio was 3:23.
Researchers said the technology could eventually help clinicians assess patients’ physical reserve and frailty before recommending treatments such as radiation, surgery or systemic therapy. However, the study did not establish whether using the model improves treatment decisions or patient outcomes.
“As the population ages, physicians need better ways to estimate physical reserve and frailty before recommending intensive treatments such as radiation, surgery or systemic therapy,” Raymond Mak, a radiation oncologist at Mass General Brigham Cancer Institute and the study’s corresponding author, said in a statement.
“Our goal is to make face photograph-based health assessment more quantitative, consistent and scalable so it can support personalised treatment decisions rather than replace physician judgment,” he said.
Also Read: How AI is transforming healthcare
The researchers initially trained FAHR-Face using more than 40 million unlabelled facial images from the WebFace42M dataset. They then adapted the foundation model for two separate tasks.
FAHR-FaceAge was trained using 743,152 facial photographs with age labels from public datasets. It achieved a mean absolute error of 5.1 years on the public APPA-REAL dataset, meaning its age estimates differed from the reference ages by an average of about five years.
The researchers used the model’s estimated age as a marker of biological ageing. However, the estimate was derived from facial appearance, not a direct measurement of cellular ageing or the physiological age of the body.
FAHR-FaceSurvival was trained on photographs of 34,363 patients from a Harvard-affiliated radiotherapy cohort. Researchers evaluated it on a separate group of 3,820 patients and tested both tools in an independent cohort of 4,892 cancer patients treated with radiotherapy in Maastricht, the Netherlands.
“The key technical advance is that FAHR-Face learns a general representation of the face from more than 40 million images, which can then be adapted to distinct health outcomes such as biological age and survival,” Hugo Aerts, co-senior author and director of the Artificial Intelligence in Medicine programme at Mass General Brigham, said.
“This foundation model approach gives us a scalable way to develop multiple facial health biomarkers, rather than training each model from scratch,” he said.
The two tools examined different facial regions and captured partly distinct information. Attention maps, which show which parts of an image most influenced the models’ predictions, indicated that FAHR-FaceAge focused mainly on the nasolabial folds, forehead and temple area. FAHR-FaceSurvival focused on regions below the eyes and around the bridge of the nose.
The researchers cautioned that these maps identify areas influencing predictions but do not establish which specific facial features or biological processes explain the results.
Also Read: Role of Artificial Intelligence in accelerating vaccine production
The researchers assessed the association between facial age and survival in a cohort of 38,183 patients treated at six Harvard-affiliated sites between 2008 and 2023. The cohort included patients who received radiotherapy for cancer and some benign conditions.
They calculated the deviation between the age estimated by FAHR-FaceAge and each patient’s chronological age. Among the patients, 33% had an estimated facial age at least five years above their chronological age, while 17% had a deviation of at least 10 years.
Patients whose estimated facial age was at least five years above their chronological age had a 21% higher hazard of death than those with a smaller deviation, after adjustment for factors including chronological age, sex, race or ethnicity, cancer site and treatment characteristics. The adjusted hazard ratio was 1.21.
Conversely, patients whose estimated facial age was at least five years below their chronological age had a 17% lower hazard of death than those whose deviation was greater than minus five years.
These findings indicated an association between estimated facial ageing and survival. They did not show that looking older causes poorer outcomes or that facial age alone can determine an individual patient’s prognosis.
The survival model showed a stronger separation between its highest- and lowest-risk groups. In the held-out test cohort of 3,820 patients, the adjusted hazard ratio for the highest-risk quartile compared with the lowest was 3.23, with a 95% confidence interval of 2.70 to 3.86.
The study also found differences in deaths during the first 90 days. In the highest score band, 64 of 122 patients, or 52.5%, died within 90 days. In the lowest score band, 20 of 975 patients, or 2.1%, died within the same period.
These were observed outcomes in groups classified retrospectively by the model. They should not be interpreted as precise individual predictions of the probability of death.
Aerts described the survival model as a further step in facial health recognition AI.
“FAHR-FaceSurvival advances the field of facial health recognition AI a step further by learning survival risk directly from facial photographs of patients with cancer,” he said.
“To our knowledge, this is the first demonstration that an AI model can directly predict cancer survival from a simple photograph,” he said.
Also Read: Using Artificial Intelligence in surgery and drug discovery
The study also examined lifestyle factors associated with facial-age estimates. Smoking showed the strongest association with higher estimated facial age, with current smokers tending to have higher deviations than former smokers, who in turn had higher deviations than people who had never smoked.
Heavy alcohol consumption and drug use were also associated with higher facial-age estimates.
The researchers said the findings suggest that facial photographs may contain multiple signals related to health. However, the observational analysis did not establish that these behaviours caused the changes identified by the model.
The researchers also tested the age-estimation model against variations in photographs, including lighting, viewing angle, facial expression and makeup. They found no statistically significant differences in estimates across several of these variations and most of the cosmetic procedures examined. Nose correction was associated with a small increase in estimated age.
“Another surprising finding was how consistent FAHR-FaceAge was across common real-world photo variations, including pose, lighting, facial expression and makeup,” Mak said. “It was also stable across most tested plastic surgery procedures, with only a small shift observed after rhinoplasty.”
The researchers nevertheless acknowledged that image quality, lighting and facial expression could affect accuracy. They said standardised photograph capture would be important if the models were used in clinical settings.
Also Read: AI, doctors’ ‘avatar’ performing surgeries boost medical tourism
The researchers said the two models could eventually complement conventional risk calculators, particularly where formal assessments of frailty are difficult to conduct. Such assessments can include frailty scales, grip-strength tests and the Timed Up and Go test, which evaluates mobility.
But the study did not test whether the models accurately identify patients who cannot tolerate aggressive treatment, nor whether using the scores improves treatment selection, survival or quality of life.
The clinical cohorts also imposed limits on how broadly the findings can be applied. They consisted of patients treated with radiotherapy, and the research was conducted mainly at centres in high-income countries. Although the authors reported performance across age, sex, race and cancer subgroups, some demographic groups in the survival-model test cohort were small.
The researchers said broader validation, stronger representation of under-represented populations, clearer governance and prospective trials are needed before the tools can be integrated into routine care. Further studies must also establish whether the models add useful information beyond existing clinical assessments and prognostic tools.
For now, the findings suggest that facial photographs might contain measurable patterns associated with ageing and cancer survival. Whether those patterns can improve clinical decisions remains a question for future research.
(Edited by Majnu Babu).
