Summary: Researchers at Mass General Brigham have developed an artificial intelligence tool called FaceAge that estimates biological age from facial photographs and uses that estimate to help predict survival outcomes in patients with cancer. In testing with more than 6,000 cancer patients, the model found that people with cancer looked, on average, roughly five years older than their chronological age, and higher FaceAge scores were associated with poorer survival.
FaceAge outperformed clinicians at predicting short-term life expectancy for patients receiving palliative radiotherapy, and clinicians’ prognoses improved when they incorporated FaceAge into their assessments. These results indicate that facial features can serve as a non-invasive biomarker for biological aging and disease risk, with potential applications in precision oncology and broader health monitoring.
Key Facts:
- FaceAge AI: Deep learning model that predicts biological age and survival from facial photos.
- Cancer Insight: Patients with cancer appeared approximately five years older than their chronological age on average.
- Clinical Impact: Providing FaceAge to clinicians improved prognostic accuracy for patients receiving palliative therapy.
Source: Mass General
Overview
A team led by Hugo Aerts, PhD, director of the Artificial Intelligence in Medicine (AIM) program at Mass General Brigham, created FaceAge: a deep learning system that estimates biological age from a simple facial photograph. The researchers show that FaceAge provides meaningful clinical information beyond chronological age and can improve survival predictions for people with cancer.
The investigators trained FaceAge on tens of thousands of face images and validated its prognostic value in multiple cancer cohorts. Higher FaceAge—meaning a face that appears older than the person’s calendar age—was linked to shorter overall survival across tumor types and cancer stages. The association remained significant after adjusting for chronological age, sex, and cancer type, particularly among patients whose FaceAge suggested they looked over 85 years old.
Estimating survival time near the end of life is challenging but crucial for treatment planning. To test clinical utility, the team asked clinicians to estimate short-term life expectancy for patients receiving palliative radiotherapy using photos and clinical context. Clinicians’ predictions were only slightly better than chance. However, when clinicians were given FaceAge estimates alongside the photos and clinical details, their accuracy improved substantially.
How FaceAge was developed and tested
FaceAge was developed using deep learning and facial recognition techniques. The model was trained on 58,851 photographs from presumed healthy individuals drawn from public datasets, and then evaluated on 6,196 patients with cancer from multiple clinical centers where facial photographs are routinely taken at radiotherapy admission. The researchers compared FaceAge scores for cancer patients against a non-cancer reference cohort and performed survival analyses to test prognostic value.
Complementary analyses included gene-based testing to explore whether FaceAge correlates with molecular markers of cellular senescence. FaceAge showed associations with senescence-related genes that chronological age did not capture, suggesting the model may reflect biological aging processes at a molecular level.
Clinical implications and future research
These findings point to a promising, low-cost, and non-invasive biomarker: a standardized AI estimate of biological age derived from a selfie or clinical photo. Potential uses include refining prognostic models, supporting end-of-life decision-making, and identifying individuals at higher risk who might benefit from targeted interventions. The authors emphasize that further validation in larger, diverse populations and rigorous ethical and regulatory review are necessary before clinical deployment.
Planned follow-up studies aim to expand testing across hospitals, examine patients at different cancer stages, monitor FaceAge trajectories over time, and assess robustness against visual alterations such as cosmetic procedures or heavy makeup.
Quotations
“We can use artificial intelligence to estimate a person’s biological age from face pictures, and our study shows that information can be clinically meaningful,” said Hugo Aerts. “A simple photo contains important signals that could help inform care plans. Individuals whose FaceAge is younger than their chronological age do significantly better after cancer therapy.”
Co-senior author Ray Mak, MD, added: “As many chronic diseases are linked to aging, accurate tools to predict an individual’s aging trajectory could enable earlier detection and intervention. With appropriate oversight, this technology could be used across many clinical applications.”
Authorship and disclosures
Additional Mass General Brigham authors include Dennis Bontempi, Osbert Zalay, Danielle S. Bitterman, Fridolin Haugg, Jack M. Qian, Hannah Roberts, Subha Perni, Vasco Prudente, Suraj Pai, Christian Guthier, Tracy Balboni, Laura Warren, Monica Krishan, and Benjamin H. Kann. Mass General Brigham has filed provisional patents on two next-generation facial health algorithms.
Funding
This research received support from the U.S. National Institutes of Health (NIH) and the European Union’s European Research Council.
About this AI, aging, and cancer research news
Author: Ryan Jaslow
Source: Mass General
Contact: Ryan Jaslow – Mass General
Image: The image is credited to Neuroscience News
Original Research: Open access. “FaceAge, a deep learning system to estimate biological age from face photographs to improve prognostication: a model development and validation study” by Hugo Aerts et al., Lancet Digital Health. DOI: 10.1016/j.landig.2025.03.002
Abstract
FaceAge, a deep learning system to estimate biological age from face photographs to improve prognostication: a model development and validation study
Background
Individuals age at different rates, and physical appearance can reflect biological aging and physiological health more accurately than chronological age alone. In clinical practice, visual assessments are subjective and vary between observers. This study aimed to develop and validate FaceAge, a reproducible deep learning method to estimate biological age from widely available face photographs.
Methods
FaceAge was trained on images from 58,851 presumed healthy adults aged 60 or older (56,304 from the IMDb–Wiki dataset for training and 2,547 from UTKFace for initial validation). Clinical utility was evaluated in 6,196 patients with cancer from cohorts in the Netherlands and the USA, including MAASTRO and Harvard thoracic and palliative cohorts. FaceAge estimates were compared with a non-cancer reference cohort of 535 individuals. Prognostic relevance was assessed using Kaplan–Meier survival curves and Cox proportional hazards models adjusted for clinical covariates, and the model’s effect on clinician predictions in palliative care was tested. A gene-based analysis explored links between FaceAge and molecular markers of senescence.
Findings
FaceAge provided independent prognostic information across cancer types and stages. Looking older correlated with worse overall survival after adjustment (per-decade hazard ratios around 1.12–1.15 across cohorts). On average, patients with cancer appeared older than the non-cancer reference group by about 4.8 years (p<0.0001). Incorporating FaceAge improved clinicians’ survival predictions for patients with incurable cancer receiving palliative treatment (AUC increased from 0.74 to 0.80; p<0.0001). Gene analysis linked FaceAge to senescence pathways not captured by chronological age.
Interpretation
A deep learning model can estimate biological age from face photographs and enhance survival prediction in cancer patients. Additional validation in larger and more diverse cohorts is necessary to confirm these findings and evaluate applicability to other diseases. With appropriate testing and regulation, FaceAge-like approaches could convert visual appearance into objective, clinically useful metrics.
Funding
Supported by the U.S. National Institutes of Health and the European Research Council.