A new experiment in AI medical applications: a single mugshot can predict cancer patients' chances of survival

The Zhitong Finance App learned that, according to reports, researchers have built an AI tool that can predict cancer patients' chances of survival through a single face photo, as a supplement to doctors' intuitive judgment, and may help guide treatment decisions when medical data is scarce.

How to do the research: A single photo gives a risk score

The study was published in The Lancet — Digital Health. Models are trained to estimate people's age; for cancer patients, they are used to estimate survival chances. The training data included more than 40 million facial images and photos of more than 30,000 cancer patients, many of whom have died. The tool will give patients a risk score: those with the highest ratings are more likely to die within 90 days.

According to the research team, these findings are the first of their kind. Doctors are generally trained to evaluate a patient's condition by observing their appearance. However, Raymond Mark, one of the study's authors and a radiation oncologist at the Bregan Cancer Institute at Massachusetts General Hospital, said that doctors “are generally not good at estimating life expectancy at death.”

“We believe that using AI, it is possible to capture part of physiological reserves and risk of death with just one facial photo.” Mark said in an interview.

The AI tool estimates age with an error range of about 5 years and can reliably classify patients into higher risk or lower risk groups; the risk of subsequent death between the two groups was different. The model works across ages, races, genders, and cancer types.

AI tools are increasingly being tested for cancer detection and treatment, from reading scanned images to help diagnose patients to analyzing genetic and lifestyle risk factors to predict patients' risk of cancer.

Expert Reminder: Clinical value has yet to be verified

Despite encouraging results, Miroslaw Bober, professor of video processing and machine learning at the University of Surrey, said the tool was not clinically ready.

“Predicting results are not the same as explaining results,” said Bob, who was not involved in the study. “The study did not show that these predictions could provide clinically useful information other than the doctor's assessment of the patient's overall health and fitness, and physician assessments were not measured in this study,” he added. Bob also said that the tool needs to be compared to the doctor's own judgment and tested in different hospitals and patient groups.

Mark and his team developed the tool. He said he was shocked by the tool's effectiveness in predicting survival. The study found that when evaluating, the algorithm mainly focuses on the bridge of the nose and the area directly below the eyes. After further investigation, Mark discovered that Hippocrates first described early signs of near death, such as sunken eye sockets, sunken temples, and a sharp nose.

According to Mark, the tool is unlikely to be used to deny or limit patients' access to medication. “We know that doctors tend to overtreat at the end of their lives.” In his opinion, this is a way to “let doctors have more information.”