BOSTON – Researchers at Mass General Brigham have developed a foundational artificial intelligence model called FAHR-Face that can estimate biological age and directly predict cancer survival outcomes from a single photograph of a patient’s face, according to a study published on Monday in the Scientific American report on research released in Lancet Digital Health. The technology is designed to transform subjective clinical evaluations into quantitative data to assist oncologists in determining appropriate treatment plans for patients.
Traditionally, physicians make visual assessments of a patient’s physical reserve, health, and frailty when entering a treatment room to decide whether intensive therapies such as surgery, radiation, or chemotherapy are appropriate. Raymond Mak, a radiation oncologist at News-Medical.net and the Mass General Brigham Cancer Institute, noted that while this clinical intuition is valuable, it remains inherently subjective.
How FAHR-Face Operates and Analyzes Facial Features
The FAHR-Face—or Foundation Artificial Intelligence for Health Recognition—model was trained by exposing an algorithm to 40 million face images without a specific assigned task to establish a general visual representation. Researchers then fine-tuned the model into two distinct tools: FAHR-FaceAge, which measures biological aging, and FAHR-FaceSurvival, which estimates cancer survival risks.
To understand the mechanics behind the model’s predictions, the research team employed a technique called attention mapping. For biological age estimation, the algorithm focused primarily on facial areas such as the forehead and the nasolabial folds, which are the creases extending from the nose to the corners of the mouth. In contrast, for cancer survival predictions, the AI concentrated on the bridge of the nose and the region directly beneath the eyes. Although the exact biological reasons why these specific facial zones correlate with survival outcomes remain unclear, the combination of both metrics yielded superior prognostic accuracy compared to either tool used independently.
Clinical Trial Results and Predictive Accuracy
Testing involving a cohort of 3,820 cancer patients demonstrated significant predictive power. Among patients flagged with the highest risk scores by the algorithm, more than half—64 out of 122—died within 90 days of their photograph being taken. A separate test group comprising more than 38,000 cancer patients similarly showed that individuals who appeared older than their actual chronological age faced elevated risks of mortality.
The algorithm proved robust against common photographic variations, including smiling, frowning, varying lighting conditions, and the presence or absence of makeup or routine plastic surgery. A notable exception was rhinoplasty, which caused the AI to rate post-surgery images slightly older than pre-surgery photographs. Furthermore, the model appeared to capture secondary health indicators, noting that heavy alcohol consumption and smoking correlated with looking older than chronological age.
Clinical Readiness and Ethical Considerations
Acknowledging the preliminary nature of the findings, Mass General Brigham has a patent pending for the technology, but the tool is not yet ready for deployment in clinical settings. The initial training dataset was drawn predominantly from a Boston-area cancer patient population consisting largely of white individuals, prompting researchers to actively expand their demographic scope to ensure equitable performance across diverse populations.
External bioethicists have also raised questions regarding the broader societal implications of facial analysis technology. Matthew Liao, director of the Center for Bioethics at New York University, cautioned that automated health assessments could introduce risks if misapplied by third parties such as insurance companies or employers, or if algorithmic errors lead to denied medical care. Despite these concerns, developers remain optimistic that quantitative facial biomarkers will eventually provide a scalable supplemental data point for personalized oncology and palliative care decisions.