AI Can Predict Cancer Survival and Biological Age, Study Finds
On October 19, 2026, researchers at Mass General Brigham published a study in The Lancet Digital Health demonstrating that an artificial intelligence model can analyze a simple facial photograph to estimate a person’s biological age and directly predict cancer survival rates, as reported by La Nación. A photograph of the face could provide information about a person’s health that is not readily apparent at a glance. Scientists state that this is the first study of its kind.
- An AI model trained on over 40 million facial images generated two algorithms to assess biological age and cancer survival directly from a photograph.
- Patients who appeared at least five years older than their chronological age faced a 21% higher risk of mortality.
- The algorithm identified specific facial regions, including the nasolabial folds and the area under the eyes, to calculate patient risk.
Development of the FAHR-Face Algorithms
The underlying system, termed Foundation Artificial Intelligence for Health Recognition (FAHR-Face), was trained on a massive database of facial images, specifically trained with more than 40 million human facial images alongside photos of over 30,000 cancer patients, many of whom have passed away. From this foundational model, the investigators developed two distinct algorithms: FAHR-FaceAge and FAHR-FaceSurvival. FAHR-FaceAge was tested on 743,152 portraits of individuals with known chronological ages. When evaluated against images outside the training set, the algorithm’s age estimations diverged from actual ages by an average of 5.1 years. According to El País, when applied to patients undergoing radiotherapy, the model revealed a mean discrepancy of approximately seven years, which researchers interpret as a marker of accelerated biological aging.
Mortality Prediction in Oncology Patients
The second tool, FAHR-FaceSurvival, prescinds of age estimation and seeks to calculate mortality risk directly from the photograph, having been trained using 34,363 facial images of cancer patients paired with clinical survival outcomes. Infobae noted that patients classified by the algorithm into the highest risk quartile experienced a mortality rate more than triple that of the lowest risk group. This stratification held true even after investigators adjusted for variables such as patient age, biological sex, cancer type, and treatment modality. Vietnam.vn reported that the tool categorizes patients into distinct risk tiers, with top-tier scores indicating heightened mortality risk within 90 days. Patients who appeared at least five years younger than their real age experienced a 17% lower risk.
Facial Regions Analyzed by the AI Models
Although both algorithms originate from the same base model, they isolate different anatomical features. La Razón detailed that FAHR-FaceAge focuses primarily on the forehead, temples, and nasolabial folds connecting the nose to the corners of the mouth. In contrast, FAHR-FaceSurvival targets the bridge of the nose and the area directly beneath the eyes. Aerts stated that the model learns survival risk directly from facial photographs, marking a novel demonstration in facial health recognition technology. After conducting further research, the team discovered that Hippocrates, “the father of western medicine,” was the first to describe sunken eyes, hollow temples, and a pinched nose as early physical signs of imminent death.
AI Helps Clinicians Evaluate Patient Fragility
Raymond Mak, a radiation oncologist at Mass General Brigham and co-author of the study, noted that physicians are trained to judge patients solely by their appearance, yet generally are not very good at estimating patient survival time in the final stages of life. The objective is to assist clinicians in evaluating patient fragility and physical reserve before recommending intensive interventions such as surgery, systemic therapies, or radiation, particularly in older populations. The analysis linked a biologically older facial appearance with habits such as smoking, heavy alcohol consumption, and drug use; smoking showed the most pronounced pattern across current smokers, former smokers, and non-smokers.
Experts Call for Validation Across Diverse Cohorts
Despite these findings, experts emphasize that the tool requires extensive external validation across diverse patient cohorts before clinical implementation. El País highlighted that the study’s cohorts originate predominantly from high-income countries, leaving certain racial groups underrepresented and sustaining demographic bias risks. Miroslaw Bober, a professor of video processing and machine learning at the University of Surrey who was not involved in the study, suggested to Vietnam.vn that the software must be rigorously tested across multiple hospitals and compared directly against physician evaluations. This investigation follows earlier work published in 2022 by a team led by Esteller, which identified genetic similarities among individuals with digital facial resemblances, though neither study establishes facial appearance as a standalone diagnostic test. Meanwhile, other research teams must still attempt to replicate the results, though in this first study it appears to indicate that the tool is quite accurate. But it remains to be known how we interpret that signal and what it contributes beyond what we already know.
Disclaimer: The information provided in this article is for educational and scientific communication purposes only and does not constitute medical advice. Always consult with a qualified healthcare provider regarding any medical condition, diagnosis, or treatment plan.
