NEW YORK (Reuters Health) - If you show people’s faces to a computer, it does better than doctors at recognizing the signs of a rare hormone disorder, researchers report in a new study.
By analyzing photos of individual faces, the computer system correctly identified seven of every 10 people in the study with acromegaly, a condition that leads to excessive growth in bones and tissues, including those of the face.
Human experts were able to correctly pick out people with acromegaly roughly six out of 10 times.
“I have to say I was surprised,” Dr. Harald Schneider, the lead author of the study and a researcher at Ludwig-Maximilians University in Germany, wrote in an email to Reuters Health. “We expected the program (at best to) achieve similar results as the experts but not to outperform them.”
People with acromegaly have too much growth hormone, which can cause a variety of problems. They usually have a large jaw and tongue, widely spaced teeth, and enlarged bones in the forehead and cheeks.
About six out of every 100,000 people have acromegaly.
In this study, the researchers took front and side facial photographs of 57 people with acromegaly and 60 people without. All the participants were Caucasian and similar in age.
The researchers wrote a computer program that analyzes facial characteristics, such as the distances between features.
The computer out-performed physicians in identifying both those who had acromegaly and those who did not.
“Apparently the computer program is better at detecting all the information that is in a face than even the experienced doctor, at least by looking at a photograph,” Schneider said.
Looking only at pictures was a disadvantage for the doctors. Schneider said the experts might have performed better if given the opportunity to do a normal examination of the patients, including their hands and feet. There was also no information available about people’s symptoms.
Schneider said he was especially encouraged by the computer’s ability to detect mild cases of acromegaly. In those cases the software was correct nearly six out of 10 times, but humans were accurate only four of 10 times.
Being able to diagnose acromegaly sooner in patients, before it develops into a more severe form, would greatly relieve suffering by giving people the opportunity for treatment earlier, Schneider said.
The study was funded by a German Research Foundation Grant, and some of the authors have received speaker fees and travel grants from pharmaceutical companies that market treatments for acromegaly.
“I see this one as yet another piece of evidence adding to the growing body of research that computer-based systems can diagnose patients as well as or better than human diagnosticians,” said Dr. James Mazoue, a philosophy professor who is the director of online programs at Wayne State University.
So far no computer-based diagnostic systems have replaced human scrutiny. Instead they are used as an adjunct.
For instance, pap smears, which are used to detect abnormal cervical cells in women, can go through a computerized optical screen, but they also receive a human analysis.
Mazoue, who was not involved in this study, believes it is morally imperative for the medical field to pursue computer-based diagnostics if they can perform better than humans.
“We need to divorce ourselves from these legacy practices that are based on human intuition. I really don’t have any doubt that in the future that’s where things are moving,” Mazoue told Reuters Health.
The acromegaly detection program, Schneider said, is not yet ready for the doctor’s office.
While the computer was good at figuring out which people had acromegaly, it was not so good at figuring out who didn’t. The computer incorrectly diagnosed acromegaly in nearly nine percent of people who did not have the disease.
The study, which is published in The Journal of Clinical Endocrinology and Metabolism, also did not include people of other races.
“It’s too small a sample to get all of the variations of acromegaly in all people,” said Dr. Randolph A. Miller, a professor of biomedical informatics at Vanderbilt University, who was not involved in this study.
Miller said he would like to see a program that could correctly identify people who don’t have acromegaly 99.9 percent of the time before it’s used in practice. But he called the system “a promising young technology.”
SOURCE: bit.ly/mKwlRb, The Journal of Clinical Endocrinology and Metabolism, online April 20, 2011.