What is FaceAge, the AI tool that can tell how healthy you are from a selfie? | Technology News

The developers of the AI tool have said that they expect to conduct a pilot study with about 50 patients starting next week. This means that FaceAge is yet to undergo proper testing before it can be deployed in hospitals to be used by doctors routinely.

What is FaceAge? What are its capabilities?

FaceAge is essentially powered by a deep learning algorithm that has been trained and developed to tell patients’ biological age from a selfie. However, the tool is designed to provide a patient’s age in health (biological age) and not in years (chronological age).

A person’s biological age is considered to be important because it could help doctors determine the most appropriate treatment for them. For example, doctors could prescribe a more aggressive treatment for a cancer patient if their biological age indicates that they are healthy enough to tolerate it.

“We found that doctors on average can predict life expectancy with an accuracy that’s only a little better than a coin flip when using a photo alone for their analysis,” Dr Raymond Mak, a radiation oncologist at Mass General Brigham and one of the co-authors of the study was quoted as saying by Washington Post.

“Some doctors would hesitate to offer cancer treatment to someone in their late 80s or 90s with the rationale that the patient may die of other causes before the cancer progresses and becomes life-threatening,” Dr Mak added.

At a press conference held last week, he recalled the case of an 86-year-old man with terminal lung cancer. “But he looked younger than 86 to me, and based on the eyeball test and a host of other factors, I decided to treat him with aggressive radiation therapy,” he said.

Four years later, Dr Mak said he used FaceAge to analyse the lung cancer patient’s face. “We found he’s more than 10 years younger than his chronological age. The patient is now 90 and still doing great,” he said.

How does FaceAge work?

Mass General Brigham researchers said that FaceAge’s training datasets comprised 9,000 photographs of people ages 60 and older who were presumed to be healthy.

A majority of the photos were downloaded from Wikipedia and IMDb, the internet movie database. The AI system was also trained using a large-scale dataset sourced from UTKFace, which comprised pictures of people between one year to 116 years old.

“It is important to know that the algorithm looks at age differently than humans do. So, for example, being bald or not, or being grey is less important in the algorithm than we actually initially thought,” Hugo Aerts, one of the co-authors of the study, said.

The study noted that no face photographs of patients and other clinical datasets were used to train the AI tool.

How accurate is the AI tool?

Researchers of the study have emphasised that FaceAge is not meant to replace but enhance a doctor’s visual assessment of a patient, otherwise known as the “eyeball test”.

The deep learning system has also undergone some testing. FaceAge was tested on photographs of over 6,200 cancer patients. These images of the patients were captured before they underwent radiotherapy treatment. The AI algorithm determined that the patients’ biological age was on average five years older than their chronological age.

The survival outlook of these patients provided by FaceAge was also dependent on how old their faces looked.

In another experiment, the researchers asked eight doctors to tell whether patients who had terminal cancer would be alive in six months. When doctors relied only on a patient’s photograph to make their prediction, they were right 61 per cent of the time. That figure rose to 73 per cent when doctors relied on the photograph as well as clinical information.

The doctors’ reached an even higher accuracy of 80 per cent when using FaceAge, along with information on medical charts.

The study also noted that an older-looking face does not necessarily lead the AI tool to predict a poor health outcome. After analysing photos of actors Paul Rudd and Wilford Brimley (when both were aged 50), FaceAge determined that Rudd’s biological age was 43 and Brimley’s was 69, as per the study. However, Brimley died in August, 2020, at 85-years-old.

Is it safe? What happens to patients’ data?

The team behind FaceAge has acknowledged that there is a long way to go before the AI tool is deployed in a real-world clinical setting as there are several risks that need to be effectively addressed.

For instance, privacy has always been a long-standing concern when it comes to AI systems that gather facial data. However, the study noted, “Our model is configured for the task of age estimation, which, in our opinion, has less embedded societal bias than the task of face recognition.”

Researchers also said that they sought to address potential racial or ethnic bias in the AI tool by quantifying “model age predictions across different ethnic groups drawn from the UTK validation dataset.” “The UTK is one of the most ethnically diverse age-labelled face image databases available publicly and, therefore, appropriate for assessing model performance in this regard, with non-White individuals comprising approximately 55% of the database,” it said.

The study also noted that FaceAge is minimally affected by ethnicity as the researchers adjusted for “ethnicity as a covariate […] in the multivariable analysis of the Harvard clinical datasets.”

Still, the developers of FaceAge have said that strong regulatory oversight and further assessments of bias in the performance of FaceAge across different populations is essential.

“This technology can do a lot of good, but it could also potentially do some harm,” said Hugo Aerts, director of the Artificial Intelligence in Medicine program at Mass General Brigham and another co-author of the study, was quoted as saying.