AI Model Predicts Cancer Survival and Mortality Risk From a Single Facial Photograph

AI Tool Predicts Health From Facial Features

Researchers at Mass General Brigham1

1

have developed an AI model trained on more than 40 million facial photographs that can predict cancer survival and estimate biological age from a single image. Published in Lancet Digital Health4

4

, the research introduces two complementary systems: FAHR-FaceAge, which estimates biological age, and FAHR-FaceSurvival, which directly predicts mortality risk in cancer patients3

3

. The technology builds on how physicians traditionally assess patient health by walking into a room and making visual judgments about frailty and overall condition, but aims to make these assessments more quantitative and consistent.

The AI model was trained on diverse datasets including more than 30,000 cancer patient photographs along with their medical outcomes3

3

. Lead researcher Dr. Raymond Mak, a radiation oncologist at Mass General Brigham, explained that the goal is to support personalized medical treatments rather than replace physician judgment. The foundation-model approach allows researchers to develop multiple health indicators from the same underlying system, making it scalable for detecting different health signals from a simple facial photograph4

4

.

Striking Accuracy in High-Risk Patients

In a test group of 3,820 cancer patients, the AI tool predicts health outcomes with notable precision. Among the 122 patients with the highest risk scores, 64 died within just 90 days of when their photograph was taken1

1

. In broader testing, patients with FAHR-FaceSurvival scores in the highest 10 percent had a 50 percent mortality rate within 90 days, compared to just 2 percent for those in the lowest 10 percent3

3

.

The AI model can predict mortality risk across many different cancer types, including those that had spread to bones or brain2

2

. In a separate test involving more than 38,000 cancer patients, those who looked older than their chronological age had a higher risk of dying1

1

. Among cancer patients treated with radiotherapy in the US, people who appeared five or more years older than their actual age had a 21 percent higher risk of dying during the average four-year follow-up, while those who looked five or more years younger had a 17 percent lower risk4

4

.

How the AI Analyzes Faces

Using attention mapping, researchers identified which facial regions each algorithm prioritizes1

1

. FAHR-FaceSurvival focuses on the area below the eyes and around the nasal bridge when predicting mortality risk3

3

. Meanwhile, FAHR-FaceAge predominantly examines nasolabial folds—the creases between nose and mouth—along with the forehead area just above eye sockets and the temples4

4

.

What remains unclear is exactly what physiological features the AI detects in these regions. Mak speculated the algorithm might identify loss of soft tissue and muscle mass around the temples, sunken eyes, and changes near the nose—features he recognizes clinically in patients nearing death. He referenced Hippocrates' writings from 400 B.C. describing "the face of death" with "a sharp nose, hollow eyes, collapsed temples."3

3

However, the AI operates purely on computer vision without access to medical literature or external data sources.

The system proved remarkably robust across real-world variations. It maintained consistency despite differences in pose, lighting, facial expression, and makeup2

2

. The AI tool predicts health while accounting for most plastic surgery procedures, though rhinoplasty caused a small shift, with the algorithm rating people slightly older after nose jobs1

1

.

Lifestyle Factors Visible in Faces

The AI model picked up on known health indicators linked to lifestyle choices. Smokers were more likely to look older than their chronological age, as were heavy alcohol drinkers1

1

. Research findings showed that heavy alcohol use, drug use, and smoking were all associated with looking biologically older in AI analysis of facial photographs4

4

. This suggests the technology can estimate biological age by detecting visible manifestations of lifestyle-related health deterioration.

Clinical Applications and Frailty Assessment

The breakthrough could assist doctors in making difficult treatment decisions for cancer patients. As populations age, physicians need better methods for frailty assessment before recommending intensive treatments such as radiation, surgery, or systemic therapy4

4

. The AI model could serve as an additional data point in clinical decision-making, helping determine which patients can tolerate aggressive interventions and which might benefit from alternative approaches1

1

.

Co-senior author Hugo Aerts emphasized that FAHR-FaceSurvival represents the first demonstration that an AI model can directly predict cancer survival from a simple photograph4

4

. The foundation-model approach provides a scalable method to develop multiple facial health biomarkers rather than training each model from scratch. Mass General Brigham has a patent pending for face health algorithms1

1

.

Limitations and Ethical Concerns

The AI model is not ready for widespread clinical use. The cancer survival model was trained primarily on a Boston-area patient population that was majority white1

1

. Researchers are working to expand their training sets to include diverse training data from different ethnic groups2

2

. The team is inviting the public to contribute photographs and receive biological age estimates to help build more representative datasets.

Ethical concerns surround any AI involvement in determining patient treatment plans. Matthew Liao, director of the Center for Bioethics at New York University, noted the tool could prove useful for identifying patients needing extra care, but warned that algorithms could misidentify patients as healthy1

1

. Professor Miroslaw Bober from the University of Surrey cautioned that "predicting an outcome is not the same as explaining it," and emphasized the tool requires testing against doctors' judgment across different hospitals and patient groups2

2

. Watch for how this technology evolves as researchers address diversity gaps and validate its performance across broader populations before it influences life-or-death medical decisions.