How an AI tool can help predict cancer survival

Previously, researchers from Mass General Brigham created an AI foundation model called FaceAge to estimate a person's biological age from a facial photograph. The model was trained on over 40 million facial images across diverse age groups and ethnic backgrounds.

The researchers later fine-tuned this model and used it to build two other models: the Foundation Artificial Intelligence for Health Recognition (FAHR)-FaceAge and FAHR-FaceSurvival models. While the FAHR-FaceAge model estimated biological age, the FAHR-FaceSurvival model predicted a cancer patient's chances of survival by evaluating facial features and other clinical information.

Using the models, researchers found that cancer patients who received radiotherapy and looked five or more years older than their actual age had a 21% higher risk of dying during the average four-year follow-up period. In comparison, patients who looked five or more years younger had a 17% lower mortality risk.

Among patients whose estimated biological age was over 20 years older than their chronological age, 21% died within 90 days. However, only 6% of patients with predicted biological ages at least 10 years younger than their chronological age died within 90 days.

In addition, 52.5% of the 122 cancer patients who were identified by the FAHR-FaceSurvival model as having a high risk of death died within 90 days of the assessment. Among the 975 patients in the lowest-risk FAHR-FaceSurvival group, only around 2% died within 90 days.

"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," said Hugo Aerts, a co-senior author of the study. "To our knowledge, this is the first demonstration that an AI model can directly predict cancer survival from a simple photograph."

Notably, the researchers found that the AI models' performance was not affected by cosmetic changes to the face, including surgery, injectables like Botox, or makeup. The study also accounted for external factors such as age, sex, race or ethnicity, cancer site, and different aspects of treatment.

Commentary

According to the researchers, the AI models are not meant to replace physicians or existing clinical evaluations, but they are an emerging tool that could potentially help inform a doctor's assessment or decision-making about patient care in the future.

"This is an extremely compelling study that stems from a large data set and its associations between facial images and biological age as well as cancer survival risk appear to be robust," said Wilbur Lam, research chair and a professor of pediatrics and biomedical engineering at Emory University. "Its true clinical utility, however, will need to be demonstrated prospectively."

Separately, Emma Pierson, an assistant professor of computer science and core faculty in the Computational Precision Health program at the University of California at Berkeley, said the study's findings were "scientifically interesting," but questioned whether physicians are already implicitly assessing patients' faces.

Before the model can be used in clinical practice, Pierson said clear evidence is needed to show it can improve predictions in a way that would not otherwise be captured.

So far, the AI models have "not shown that these predictions add clinically useful information beyond doctors' assessments of a patient's general health and fitness, which were not measured in the study," said Miroslaw Bober, professor of video processing and machine intelligence at the University of Surrey. "This is a promising proof of concept, but it is not ready for clinical use. The tool still needs testing against doctors' own judgement and across different hospitals and patient groups."

Pierson also raised concerns about the potential for racial bias in the AI models. "Any time you train something on someone's face, you are implicitly including patient race in the model's predictions," she said, adding that using race in medical predictions is "enormously controversial."

Overall, Suchi Saria, a professor and director of the Machine Learning and Healthcare Lab at Johns Hopkins University, said the study suggests that "there is real prognostic information encoded in facial appearance … To me, that's the important scientific contribution: It suggests there is measurable health information in the face that we haven't previously been able to quantify very well."