Researchers developed a foundation model, called Foundation AI for Health Recognition-Face (FAHR-Face), that estimates biological age and predicts survival from routine facial photographs of patients with cancer. Study findings were published in The Lancet Digital Health.

“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,” said co-corresponding author Raymond H. Mak, MD, a radiation oncologist at Mass General Brigham Cancer Institute and a faculty member in the AI in Medicine (AIM) program at Mass General Brigham. “Our goal is to make face photograph–based health assessment more quantitative, consistent, and scalable so it can support personalized treatment decisions rather than replace physician judgment.”

The two tasks of the foundation model could result in clinical biomarkers for improving prognostic accuracy and clinical decision-making, especially in settings where traditional biomarkers or comprehensive assessments are limited.

Study and Model Methods

The foundation model was trained on more than 40 million facial photographs using a masked autoencoder, where random patches were masked and the pixels were reconstructed. From there, the model was fine-tuned to estimate biological age (FAHR-FaceAge) and survival prediction (FAHR-FaceSurvival). The model built on prior research from the same team looking at using facial images to estimate biological age to improve prognostication.

For the biological age task, FAHR-FaceAge underwent two-stage, age-balanced fine-tuning on 10 public image data sets of 743,152 facial photographs. The model was then evaluated on a public data set and an independent patient data set of 38,183 patients who received radiotherapy at Harvard. For the survival prediction task, FAHR-FaceSurvival was fine-tuned on 34,363 photographs of patients with cancer. During development, the researchers focused on testing the robustness and independence of the model. The models were also tested on two independent data sets of patients with cancer.

“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,” said co-senior author Hugo J.W.L. Aerts, PhD, Director of AIM at Mass General Brigham. “This foundation model approach gives us a scalable way to develop multiple facial health biomarkers, rather than training each model from scratch.”

The model was externally validated on a data set of 4,892 patients with cancer who were treated with radiotherapy in the Netherlands.

Key Findings

In terms of biological age estimation, FAHR-FaceAge showed the lowest mean absolute error on public data sets compared with benchmark models. Accuracy was maintained across the full human lifespan.

FAHR-FaceAge outperformed FaceAge, from the team's prior research, in terms of survival prognostication, with an area under the curve of 0.66 for FAHR-FaceAge at 3 months, 0.64 for FaceAge, and 0.63 for chronological age. These trends continued at 6 months, 1 year, and 2 years.

FAHR-FaceSurvival demonstrated robust prediction of cancer mortality. Patients in the highest-risk quartile showed a significantly increased risk of mortality than patients in the lowest-risk quartile (adjusted hazard ratio [HR] = 3.23; P < .0001).

“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,” Dr. Aerts said. “To our knowledge, this is the first demonstration that an AI model can directly predict cancer survival from a simple photograph.”

Both models showed generalizability in independent validation across age, sex, race, and cancer subgroups. FAHR-FaceSurvival demonstrated an overall concordance index (C-index) of 0.72. By race, the C-index ranged from 0.63 to 0.75, and between 0.62 and 0.75 by age group. In terms of cancer types, the lowest C-index was 0.53 for Hodgkin lymphoma and the highest was 0.88 for HPV-positive oropharyngeal cancer.

Both models’ predictions were only weakly correlated with each other and were independently associated with survival when analyzed in tandem. They also provided complementary prognostic information for patients with cancer based on distinct facial regions.

Lifestyle factors contributed greatly to signs of facial aging with smoking showing the strongest association with looking biologically older in facial images.

“Another surprising finding was how consistent FAHR-FaceAge was across common real-world photo variations, including pose, lighting, facial expression and makeup,” noted Dr. Mak. “It was also stable across most tested plastic surgery procedures, with only a small shift observed after rhinoplasty.”

The study authors believed that routine quantitative facial analysis could enhance risk stratification for patients with cancer when traditional biomarkers are unavailable.

Going forward, the team is expanding their research through the FaceAge Research Study open to the public.

DISCLOSURES: This study was funded by the E.U. European Research Council. Dr. Mak reported being on an Advisory Board for ViewRay and AstraZeneca; Consulting for AstraZeneca, Varian Medical Systems, and Sio Capital Management; Honorarium from Novartis and Springer Nature; and Research Funding from the National Institute of Health, ViewRay, AstraZeneca, Siemens Medical Solutions USA, Inc., and Varian Medical Systems. Dr. Bitterman, who is also on the board for ASCO AI in Oncology, reported Editorial, unrelated to this work: Associate Editor of Radiation Oncology of HemOnc.org, Associate Editor, Annals of Oncology, and Associate Editor, JCO Clinical Cancer Informatics; Research funding, unrelated to this work: American Cancer Society/American Society for Radiation Oncology and National Institutes of Health/National Cancer Institute; Advisory and consulting, unrelated to this work: MercuriaAI. Dr. Aerts reported advising and consulting for Onc.AI, Love Health, Sphera, Editas, AstraZeneca, and Bristol Myers Squibb, unrelated to this work. Dr. Kann reported research funding from the Botha-Chan Low Grade Glioma Consortium. Drs. Haugg, Mak, and Aerts reported that Mass General Brigham has filed a U.S. provisional patent application on next-generation facial health algorithms. All other authors declared no competing interests.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.