NEW YORK – Researchers at Children's National Hospital in Washington, D.C., have developed a new biometric analysis tool that can screen children for 128 genetic diseases just by looking at their facial features.
While the tool hasn't yet been approved for patient care, it was licensed by MGeneRx to seek regulatory approval and commercialization for its use as a diagnostic and screening tool.
The machine learning model stratifies patients based on their risk of genetic conditions, said Marius Linguraru, principal investigator in the Sheikh Zayed Institute for Pediatric Surgical Innovation at the hospital. It was developed by looking at photos of patients of all races, ages, and genders with genetic conditions to see how those conditions manifested on the face.
It is intended to be used as a population screening tool that can indicate a patient who might benefit from further referral to a genetic counselor, he said.
The tool, which has been in development since 2012, was trained with both pictures of children without a genetic syndrome to learn what variations in facial characteristics are normal and expected, as well as pictures of patients with diagnosed genetic disorders to identify key markers, he said.
It was trained with 2,800 children from 28 countries who had 128 different genetic conditions, including Williams-Beuren syndrome, Cornelia de Lange syndrome, and Down syndrome, he said, from newborns to patients in late adolescence. The control group consisted of 1,400 pictures. The pictures were retrospectively pulled from three publicly available databases, as well as archives from Children's National and pictures taken on a smartphone at the hospital.
The deep learning architecture is structured into three neural networks: Network A, which standardizes the photos; Network B, which detects the facial morphology; and Network C, which performs the genetic syndrome risk estimation.
The team grouped the photos into smaller subsections for cross validation and retrained the tool, which Linguraru said showed improvements. Those improvements weren't surprising, because as the machine learning algorithm sees more patients, it "becomes better at determining" the relevant facial characteristics.
"The more data it has, the smarter it becomes, and the more focused it is," he said.
Currently, the tool can identify patients who may have genetic disorders with 88 percent accuracy, 90 percent sensitivity, and 86 percent specificity for any genetic conditions, according to a study from the researchers published in The Lancet Digital Health earlier this month.
The machine learning model performs "deep analysis of the patterns of the face," Linguraru said. The algorithm looks at "everything on the face," such as patterns of lines and indentations on the face, the distance between the eyes, and the length of the philtrum. It's "looking for changes" in those measurements, he said.
It then measures those distances and analyzes the patterns of those lines, returning a result in about one second that lays out the patient's risk of a genetic disorder. Although it can be used on people of all ages, Linguraru said the focus of development was for use with pediatric patients under 21 years old.
By looking at a global population that included people of all races, and those who were mixed-race, Linguraru said the team has worked to bypass potential biases — there have been issues in other facial recognition software tools that have been less effective in detecting genetic syndromes in non-Europeans.
The team tried to have "as much data [that's] as diverse as possible" to ensure the tool would be able to pick out important facial characteristics on all patients, he said. There are also differences between the patterns and measurements in different ethnic groups that the tool had to be trained to be aware of, he said.
However, the researchers noted in the Lancet study that the tool was more accurate in white and Hispanic populations than African and Asian populations — 90 percent and 91 percent accuracy in white and Hispanic patients, versus 84 percent and 82 percent accuracy in African and Asian populations, respectively.
They also said that they used a broad definition of racial or ethnic categories that may not distinguish between the finer variations within populations — for example, the Asian group included patients from China, Japan, India, and Thailand. "This broad grouping of individuals might be another reason for the difference in performance," they wrote.
Accuracy was similar in male and female children, and slightly better in children between 2 and 5 years old, the study found.
The next step for the tool is clinical validation using a prospective patient cohort and determining "an adequate operating threshold" in the classifier that provides a "clinically meaningful" sensitivity and keeps an acceptable specificity, the researchers wrote.
Linguraru was careful to note that the tool isn't intended to be used by a geneticist. Instead, it is for "that first level of care," such as a primary care doctor, who can direct the patient further.
Providing earlier detection of genetic disorders is beneficial because it significantly improves the survivability rates, as patients are able to see genetic specialists sooner and receive preventive care, Linguraru said.
Other researchers have used biometric analysis for medical diagnosis, including a German team that published a paper in 2003 investigating whether a computer can recognize disease-specific facial patterns, a University of Oxford team, and researchers in China who developed a tool for recognizing facial features of Turner syndrome.
In addition, Boston-based FDNA has commercialized facial recognition software for use in genetics, Face2Gene, which in a 2015 study published in Clinical Genetics had 87 percent accuracy.
However, a more recent study from earlier this year in the American Journal of Medical Genetics found an overall diagnostic yield of 57 percent, which increased to 82 percent when cases diagnosed with syndromes not recognized by Face2Gene were removed.
All of the intellectual property relating to Children's National Hospital's tool has been licensed to MGeneRx, which will develop it into a commercial product.
Nasser Hassan, the acting CEO of MGeneRx, said that the firm plans to utilize the technology in different ways, including as a smartphone app, because it would be the "most direct route to get to a pediatrician" and other healthcare workers. It also plans to work with different hospitals to integrate the tool into their electronic medical records, allowing pediatricians to use it directly in their offices for screening with other medical information, he added.
The primary focus right now is on using it as a screening tool, and that's what it is working on, in addition to working with the US Food and Drug Administration to further develop the tool. MGeneRx also sees potential for the technology as a diagnostic tool for rare diseases — something it is actively exploring, Hassan said.
MGeneRx is currently fundraising to further develop the tool and move toward commercialization and is looking for government, private, and philanthropic funding, he said. The company is also refining the tool so that it will be fully functional in all different mobile devices, he added.
It would also likely be licensed by a hospital or health system for use, rather than as a direct-to-consumer app — although it could conceivably be used directly. Hassan noted that for such a "sensitive subject," patients would want to go through medical practitioners to receive appropriate advice.
The company is working to commercialize it in the US first, hopefully within a year, then expand to other countries, including low- and middle-income countries where children may not have access to genetic counseling, he said. It would be available wherever there's access to the internet and adapted to the environment, he continued.
"We're really developing it because of the social impact of a product like this," Hassan said.