Cancer, study in Lancet: a selfie and AI to estimate the probability of survival

How artificial intelligence works

The researchers have developed an artificial intelligence model called FAHR-Face (Foundation Artificial Intelligence for Health Recognition-Face), trained on over 40 million images of faces of people of different ages and ethnic origins. Starting from this system, two models with different functions were created: FAHR-FaceAge, designed to estimate biological age through a photograph, and FAHR-FaceSurvival, which combines the analysis of facial features with other clinical information to evaluate the survival chances of cancer patients. "The age estimation algorithm examines the area around the temples, the forehead, the area above the eyes and the lines that connect the nose to the corners of the mouth," study author Raymond Mak explained to CNN. “Instead, the algorithm that assesses mortality risk focuses on completely different areas of the face: the bridge of the nose and the area under the eyes.”

The results of the study on cancer patients

To evaluate the reliability of the FAHR-FaceSurvival model, the researchers compared the mortality risk estimates developed by artificial intelligence with the clinical data of cancer patients, verifying how many had actually died in the 90 days following the evaluation. The comparison showed that, among the 122 patients for whom the system had indicated a high mortality risk, more than half, equal to 52.5%, died within 90 days. Among the 975 patients classified in the lowest risk group, the percentage of deaths in the same period was approximately 2%. The research team also examined the results of the FAHR-FaceAge model. Among the 1,060 patients to whom the system had attributed a biological age more than 20 years higher than their chronological age, 21% died within 90 days. The percentage was instead approximately 6% among the 4,195 patients with an estimated biological age at least 10 years lower than their chronological age.

Applications and limitations of the study

As explained by the authors, the next step will be to verify whether facial analysis can become an additional tool to early identify patients who may need further investigations and help specialists in therapeutic decisions. “If we could detect signs of accelerated aging early, we could use this data for screening or to identify patients who would benefit from further testing for serious conditions,” Mak said. However, the technology requires further checks before any use in medical practice. "This is an extremely interesting study, based on a large data set. The correlations identified between facial features, biological age and the risk of cancer survival appear solid. However, its real clinical utility will have to be demonstrated through prospective studies," said Wilbur Lam, professor of pediatrics and biomedical engineering at Emory University, who was not involved in the study. Miroslaw Bober, professor at the University of Surrey, also urged caution. “The authors link the regions of the face used by AI to what doctors call the 'dying look.' The study, however, did not test whether AI can detect elements that a doctor would miss,” Bober noted.