Researchers at Mass General Brigham developed an AI model that sorted cancer patients into groups with substantially different death hazards using face photographs, although its clinical benefit remains unproven.
The study appeared in The Lancet Digital Health on Oct. 19. The prepublication author manuscript reported that patients with FAHR-FaceSurvival scores of at least 0.75 had 3.22 times the adjusted death hazard of those with scores below 0.25. The 95% confidence interval was 2.69 to 3.85. The finding does not show that using the model changes treatment outcomes.
The manuscript says the survival model was trained on 34,389 photos of cancer patients and tested on a separate group of 3,822 patients. It was also validated in 4,892 patients at the Maastro radiotherapy clinic.
FAHR-FaceSurvival is one of two tools derived from FAHR-Face, which Mass General Brigham said was pretrained on more than 40 million facial images. The other tool, FAHR-FaceAge, produces a model-based estimate of biological age from a photograph rather than predicting survival risk directly.
The age model showed a separate association with mortality. In the manuscript's 38,211-patient Harvard radiotherapy cohort, patients assigned a facial age at least five years above their chronological age had a 21% higher adjusted death hazard than patients assigned a difference below five years. That comparison group included patients whose estimated facial age was younger than their chronological age. Median follow-up was 2.1 years.
Raymond Mak, a radiation oncologist at Mass General Brigham Cancer Institute, said the facial assessments could give doctors another consideration when weighing intensive treatment, while leaving treatment decisions to clinicians. El País reported that Leonor Cerdá Alberich, a biomedical-imaging researcher at La Fe Health Research Institute, said researchers have yet to establish why facial age tracks mortality or whether it adds useful information beyond standard clinical measures.
The cancer cohorts came from radiotherapy centers, mainly in Western countries. The authors raised concerns about protecting facial images and called for validation in more diverse populations and prospective trials to determine whether the tools improve patient care.
