SALT LAKE CITY — For parents of children with undiagnosed genetic disorders, it can be extraordinarily painful not knowing the reason for a child's symptoms — delayed speech, seizures, heart defects or extra fingers, according to clinical geneticist Dr. Karen Gripp.
Now, new artificial intelligence technology can identify genetic disorders using just a photo of a patient's face, potentially bringing peace of mind to some parents of children with rare disorders. A new study published this month in the peer-reviewed medical journal Nature Medicine found that an algorithm was better than clinicians at accurately identifying genetic syndromes with recognizable facial features. The algorithm powers an app called Face2Gene.
In one test, the AI program achieved 96.88 percent accuracy when identifying patients with Cornelia de Lange Syndrome, a developmental disorder characterized by facial features including thick, arched eyebrows, low-set ears, and a small, upturned nose. By comparison, 65 experts examining the same set of photos had a 75 percent accuracy rate. In a clinical-setting experiment involving 502 images of people with different disorders, Face2Gene correctly identified the correct syndrome in its top 10 suggestions 91 percent of the time.
Eight percent of the population is affected by syndromic genetic conditions, according to the study, which was conducted by artificial intelligence and precision medicine company FDNA. While some conditions are invisible to the eye, about 30 to 40 percent come with alterations in the face and skull, a 2014 University of Oxford study said. For example, Down syndrome is characterized by a round face, almond-shaped eyes and a protruding tongue.
The study comes at a time when applications for artificial intelligence in medicine are rapidly expanding. Research suggests AI is better than doctors at detecting lesions likely to develop into cervical cancer, flagging abnormal chest X-rays and even predicting Alzheimer's using brain imaging.
The technology behind Face2Gene program was created using a data set of 17,000 photos of people diagnosed with over 200 distinct genetic syndromes, according to the study. The program crops an image of the patient's face into different regions, assessing the probability that each part of the face matches each syndrome, and then aggregates those results to see which syndrome is the best fit.
Today, the app is being utilized by thousands of clinical geneticists around the world, according to Gripp. Each time a doctor uploads a new photo and inputs additional information about a patient, the program's suggestions become more accurate. In a little over a year, the program has gone from being able to identify about 200 conditions to more than 1,000, she added.
"It’s using existing AI technologies on a problem they haven’t been used for before," said David Forsyth, professor of computer science at University of Illinois at Urbana–Champaign, whose primary research area is computer vision. Face classifier programs have been used in the past to help Facebook users know who to tag in a photo, for example, but this is the first time Forsyth has seen the technology applied in the medical field, he said.
Gripp, chief of the division of medical genetics at the Alfred I. duPont Hospital for Children in Delaware and chief medical officer for FDNA, said the program is a valuable tool in the hands of trained clinicians who can evaluate contextual information, order additional testing and ultimately provide a diagnosis.
For example, Gripp used Face2Gene to diagnose a young boy with Smith-Lemli-Opitz syndrome, a genetic disorder that results in an enzyme deficiency and affects growth. The boy's family came to Gripp "desperate for a diagnosis," she said. Their child was born underweight, required several surgeries and was experiencing delayed development. "But when you looked at him, his facial features did not look particularly unusual," Gripp said.
The doctor took a picture of the boy on her phone and used Face2Gene to evaluate facial clues in real time. The program, able to detect subtle characteristics that were not immediately apparent to Gripp, suggested Smith-Lemli-Opitz syndrome was the closest match.
"We were able to do much more targeted testing and ultimately diagnose him in that way," Gripp said.
The program saved Gripp time and guesswork, and more importantly, provided the family with the answer they were searching for, she said.
According to one European survey, 40 percent of rare diseases are initially misdiagnosed. Accurate, early diagnosis for children with certain genetic conditions can make a difference. For example, some therapies or treatments, like enzyme replacement, are more effective the earlier a kid starts, said Gripp. In addition, misdiagnosis by a clinician can affect a family's decision to have more children based on misinformation about the recurrence risk.
But according to Gripp, lack of diagnosis is a much more common problem than misdiagnosis.
"Not having a specific diagnosis can be very unsettling," she said. "Moms think about what happened during the pregnancy and might feel guilty thinking something they did or did not do caused the disorder," Gripp said.
The AI technology is not only being used for children but also adults who have grown up without a diagnosis.
Christoffer Nellåker, a research fellow at the University of Oxford and expert in rare diseases, told New Scientist, “The real value here is that for some of these ultra-rare diseases, the process of diagnosis can be many, many years. This kind of technology can help narrow down the search space and then be verified through checking genetic markers,” he said. “For some diseases, it will cut down the time to diagnosis drastically.”
Because of the complexities of genetic diagnoses, Gripp doesn't imagine jobs for doctors in this field will ever become obsolete, despite the fact that AI surpasses human ability to perform specific tasks.
Dr. Lloyd Minor, professor of otolaryngology, head and neck surgery, at Stanford University, wrote he started to sense anxiety about job security among researchers in 2017, when Stanford Medicine announced it had developed an algorithm that could interpret chest X-ray images to diagnose more than a dozen medical conditions.
"Without a doubt, algorithms will play a vital and growing role in health care. However, while they may soon become superhuman at performing certain tasks, algorithms do not have the general intelligence that people do," Minor wrote for Stanford Medicine's blog, Scope.
"Somebody needs to examine the patient, interpret results and talk to the family," said Gripp. "You could never take away the element of human interaction."2 comments on this story
In fact, as a clinical geneticist for the past 20 years, Gripp has seen the field change dramatically. But as technologies have improved, work hasn't slowed down for clinicians. The opposite has occurred.
"Now that there are more things to consider, you need more geneticists and genetic counselors to explain what is going on and help people make choices based on the risks," Gripp said. "I think the way we do our job will continue to change dramatically, for the better, with the help of AI tools."