Dr. Maximilian Muenke has a superpower: He can diagnose disease just by looking at a person’s face.
Specifically, he can spot certain genetic disorders that make telltale impressions on facial features.
“Once you’ve done it for a certain amount of years, you walk into a room and it’s like, oh, that child has Williams syndrome,” he said, referring to a genetic disorder that can affect a person’s cognitive abilities and heart.
And that’s an incredibly useful skill, even as genetic sequencing becomes more widespread. For one thing, it can be the factor that sends someone to get a genetic test in the first place. For another, people in many parts of the world don’t have access to genetic tests at all.
That’s inspired years of effort to train a computer to do the same thing. Software that analyzes a patient’s face for signs of disease could help clinicians better diagnose and treat people with genetic syndromes.
Some older attempts at facial analysis relied on large, clunky scanners — a tool better suited to a lab, not the field. Now, in the era of smartphones, such efforts have a whole new promise. Face2Gene, a program developed by Boston-based startup FDNA, has a mobile app that clinicians can use to snap photos of their patients and get a list of syndromes they might have.
Meanwhile, Muenke and his colleagues at the NIH last month published an important advance: the ability to diagnose disease in a non-Caucasian face.
It’s a promising preliminary sign. But if facial recognition software is to be widely useful for diagnoses, software developers and geneticists will need to work together to overcome genetics’ systemic blind spots.
Diagnoses vs. probabilities
The algorithms in general work on the same principles: measuring the size of facial features and their placement to detect patterns. They’re both trained on databases of photographs doctors take of their patients. The NIH works with partners around the world to collect their photos; FDNA accepts photos uploaded to Face2Gene.
But they differ in a key way: Whereas the algorithm the NIH uses can predict if someone has a given genetic disorder, the Face2Gene algorithm spits out not diagnoses, but probabilities. The app describes photos as being a certain percent similar to photos of people with one of the 2,000 disorders for which Face2Gene has image data, based on the overall “look” of the face as well as the presence of certain features. However, the app won’t give clinicians a yes or no answer to the question of, “Does my patient have a genetic disorder?”
That’s intentional. Face2Gene is meant to be more like a search engine for diseases — a means to an end.
“We are not a diagnostic tool, and we will never be a diagnostic tool,” said FDNA CEO Dekel Gelbman.
Drawing that bright line between Face2Gene and “a diagnostic tool” allows FDNA to stay compliant with FDA regulations governing mobile medical apps while avoiding some of the regulatory burden associated with smartphone-based diagnostic tools.
The algorithm the NIH uses — developed by scientists at Children’s National Health System in Washington, D.C., — seems to work pretty well so far: In 129 cases of Down syndrome, it accurately detected the disorder 94 percent of the time. For DiGeorge syndrome, the numbers were even higher: It had a 95 percent accuracy rate across all 156 cases.
Face2Gene declined to provide similar numbers for their technology. “Since Face2Gene is a search and reference informational tool, the terms sensitivity and specificity are difficult to apply to our output,” Gelbman cautioned.
But there’s one big stumbling block for both of them, a problem that has dogged medical genetics for decades: Data for non-white populations is sorely lacking.
“In every single textbook, the ones we had [when I trained] in Germany and the major textbooks here in the US, there are photos of individuals of northern European descent,” Muenke said. “When I told this to my boss, he said there have to be atlases for children from diverse backgrounds. And there aren’t. There just aren’t.” (Today there is that resource, based on Muenke and the NIH’s work.)
So diagnosing diseases from a face alone presents an additional challenge in countries where the majority of the population isn’t of northern European descent, because some facial areas that vary with ethnic background can often overlap with areas that signify a genetic disorder. Eventually, the software will also have to be able to tackle people with mixed ethnic backgrounds, too. “We have thought about it but haven’t gone there yet,” Muenke said.
For example, children with Down syndrome often have flat nasal bridges — as do typically developing African or African-American children. Across different races and ethnicities of children there were only two reliable identifiers that could be used to diagnose Down syndrome — the angles between landmark points on the child’s nose and eye, according to a paper Muenke and Marius Linguraru at Children’s National published with their colleagues earlier this year. All of the other “typical” features weren’t significantly more likely to show up when children were compared to ethnically matched controls.
In fact, using a Caucasian face as a reference can sometimes be the least representative choice. “One of the findings that I’m very interested in [in] our recent study was that the population that we found to be most different from the others, in terms of facial patterns characteristic of DiGeorge syndrome, was the Caucasian population,” Linguraru said.
To continue to fix this problem, both the NIH and Face2Gene need help from more researchers who can upload more patients’ faces — but that’s easier said than done. Confirming a suspected disorder with genetic tests is standard practice today, and there are no genetic labs based in Africa registered in the NIH’s Genetic Testing Registry. Asia and South America are also relatively underserved.
Those numbers also reflect the general patterns of distribution for medical geneticists. “Most practitioners are located in North America and Europe,” Gelbman said. Nigeria, for example, doesn’t have a single medical geneticist in the entire country.
It’s possible that might change, with time and effort. In addition to his work as a researcher, Muenke directs a program that brings health care professionals from developing countries to the US for a month-long crash course in medical genetics. (The program is funded by the NIH’s Fogarty International Center; President Trump eliminated funding for the center in his 2018 “skinny” budget proposal announced in March.)
For now, both algorithms have shown that they can handle a diverse patient set. FDNA scientists published a paper in January showing that their algorithm could better identify Down syndrome after being trained with a more diverse set of faces, and Muenke and Linguraru have also published papers this year demonstrating their algorithm’s ability to identify genetic disorders correctly in children across a variety of ethnic backgrounds.
As both groups work on recruiting more researchers, they are also working to push their tech forward. FDNA is working on establishing partnerships with pharmaceutical companies to start their commercial outreach. In theory, these partnerships could contribute to precision medicine efforts or help companies develop new therapies for rare diseases.
Meanwhile, Linguraru has his eyes on eventual FDA approval for the algorithm the NIH has used. The ultimate goal would be a simple tool that any doctor could use anywhere to get fast results and better diagnose their patients.