Andrew was playing under the summer sun in the backyard. As the four-year-old’s parents watched, they noticed something seemed off. Perhaps it was his unusually small head or the after-effects of the surgery to correct his congenital disorder.
When Andrew’s parents consulted Dr. Karen Gripp, Professor of Pediatrics at Nemours Children’s Hospital, she decided to investigate. In addition to conventional procedures, she ran a quick diagnosis on Face2Gene, a computer vision-powered app that looks for indications of rare diseases. The facial picture uploaded to the app showed a strong match for Smith-Lemli-Opitz syndrome (SLO), a rare disease that affects about 1 in 40,000 children.
“The family was under the impression that this condition had been ruled out,” says Dr. Gripp. Subsequent tests confirmed the genetic diagnosis. Andrew’s facial features suggested a mild form of SLO. Although this came as a shock to his parents, Andrew was swiftly admitted to a metabolic disease clinic and put on the appropriate nutrition and medication.
“The family is thankful that they have a clear diagnosis that explains his ongoing behavioral and learning challenges,” adds Dr. Gripp. They are now aware that SLO has a 25% recurrence risk in the family.
There are over 10,000 rare diseases, and 75% of them affect children. About one-third of these children won’t live to see their fifth birthday. Rare diseases don’t draw much attention, and their detection is painfully challenging.
In Andrew’s case, the breakthrough came with the rapid diagnosis, thanks to the Artificial Intelligence-powered (AI) computer vision app. Are facial pictures credible indicators of these deadly conditions? Can AI detect rare diseases when so little information is available to train the algorithms? This article will address these questions and share how physicians adopt such solutions today.
<Andrew is an assumed name to protect the child’s identity>
How rare diseases hide in plain sight
Two factors make rare genetic diseases a lot deadlier—poor awareness and a paucity of targeted treatments.
One in 12 babies is born with a rare disease. Despite over 300 million rare disease patients worldwide, there is low awareness about these conditions. With over 10,000 rare diseases affecting children, the patient count is spread thin with an extremely long tail. Unlike Andrew, the conditions of most kids go undetected for years until symptoms become severe. To complicate matters, access to geneticists isn’t easy. An appointment could take months.
Since pharma companies prioritize drug discovery for diseases that affect sizeable populations, most rare conditions lack a cure. With just a few thousand patients affected by each rare disease, research efforts aren’t commercially viable. Drug discovery for rare diseases is relegated to the sidelines despite incentives to treat orphan diseases.
Technology is stepping up to solve the twin challenges of making a correct diagnosis and offering high-quality care.
Visually detecting rare diseases with AI
“Can you identify someone with Down’s syndrome?” asks Moti Shniberg, a serial entrepreneur specializing in computer vision. Down’s syndrome is a genetic disorder that causes intellectual disability and developmental delays in children. This condition manifests with distinct facial conditions and is easily identifiable.
Since this condition affects one in 700 children, “chances are that we have come across someone with the disease,” adds Shniberg. After the first few times we see children living with this condition, we tend to identify such cases instantly. Our brain gets trained to detect the syndrome visually without any specialized medical expertise.
However, extending this visual expertise to other rare diseases is a challenge, even for experts. With thousands of rare diseases with wildly different indications and each one affecting just a few thousand children globally, there aren’t enough examples for our brains to learn.
This is where computer vision can help. Shniberg built Face.com, a facial recognition platform that discovered over 18 billion faces on the internet. He sold the company to Facebook in 2012. When he and his partner, Lior Wolf, looked at other impactful problems to solve with computer vision and deep learning, they stumbled upon rare genetic diseases.
They founded FDNA, a company that develops AI-driven applications such as Face2Gene to detect rare genetic diseases visually. Their algorithms are engineered to learn from just a handful of historical patient photos per rare disease.
Over the past ten years, FDNA has built a database covering 5,000 rare diseases by working with geneticists. “1500 of these diseases can be detected with our facial recognition algorithms,” explains Shniberg. “The other 3,500 diseases leverage clinical-feature analysis that includes natural language processing.” This database powers Face2Gene and helps diagnose conditions with a single facial photograph taken by smartphones.
How does AI visual detection compare with genetic testing?
In 2003, scientists sequenced the entire human genome. This milestone led to much hype around how genotyping could detect and treat genetic diseases. About 20 years later, companies can sequence the complete DNA of individuals for a few hundred dollars. However, these advances haven’t transformed the detection of genetic diseases yet.
Spotting diseases using genotyping is like searching the ocean for a lost boat. “We are talking about a large amount of data since our genetic material is very complex,” adds Dr. Gripp. This approach isn’t very efficient on its own—at least not yet.
Over the same period, we have seen remarkable progress with phenotyping, which uses observable characteristics to understand living things. The study of facial features to detect underlying diseases falls into this discipline. Thanks to new AI and computer vision capabilities, AI-driven algorithms can detect diseases instantly. They can do this through visual inspection of medical images and using inputs such as human voice, typing patterns, and other digital biomarkers.
In practice, the two approaches of genotyping and phenotyping aren’t competitive but complementary. Facial phenotyping can act as an initial diagnostic tool by helping us arrive at a rapid and efficient shortlist.
“If we know what genes we need to look at because we have already assessed the patient, we have a pretty good differential diagnosis,” clarifies Dr. Gripp. “Then, it’s much more straightforward to analyze the data that we can get from genetic testing.”
How accurate are these solutions?
Facial detection algorithms are good but not perfect. It can be costly to clear patients incorrectly when they have an underlying serious condition, and it is also harmful to trigger false alerts in healthy individuals.
As per FDNA’s publication in Nature Medicine, their solution reports a 91% accuracy for the top ten predictions: in 91% of cases, the actual disease appears in the algorithm’s top ten recommended conditions. At this point, AI needs some human assistance.
“This is a tool that makes suggestions, and we shouldn’t expect it to make a diagnosis,” clarifies Dr. Gripp. “We use this as an additional tool to assess the patient,” she adds. When such a tool is paired with a practitioner, it can be a powerful clinical aid.
Are there areas where this approach could fall flat? “This solution cannot address purely neurodegenerative conditions,” shares Dr. Gripp. Computer vision is of limited use when conditions don’t involve changes in the physical body structure. These are areas where other biomarkers and clinical feature analysis come in handy. Such combinatorial approaches are essential to cover the entire range of 10,000 rare diseases.
Algorithmic bias is another risk to address. “Rare diseases are likely to have higher rates of diagnosis in populations with better access to health care,” says Dr. Jennifer Kraschnewski, Professor of Medicine at the Penn State College of Medicine. “This could result in an unintentionally biased data set for AI training, with downstream effects on poorer identification of rare diseases in underrepresented populations.”
The state of clinical adoption
Shniberg shares that FDNA worked with over 5,000 institutions across 150 countries last year. “Today, about 70% of geneticists worldwide use our tool,” he adds. The jump in the adoption of telemedicine during the pandemic has accelerated FDNA’s success. Their app can be readily integrated into telemedicine sessions to share live recommendations to practitioners based on patients’ facial features.
The FDNA team is working to improve adoption by taking the tool to pediatricians, partnering with telemedicine firms, and providing the app directly to parents. The company has an ambitious goal of expanding the coverage from 80,000 patients in 2021 to over one million in the next two years.
Achieving the adoption of AI tools requires a multi-pronged approach. “Medical schools and continuing medical education courses must integrate such tools into practice to ensure that we appropriately leverage these applications while caring for our patients,” adds Dr. Kraschnewski.
A future of accessible treatments for rare genetic diseases
Despite the FDA’s financial incentives to treat rare genetic diseases, just 5% of these 10,000 conditions have a treatment. Shniberg believes their goal of reaching one million rare disease patients can spark a chain reaction to address the inherent challenges in this space.
Firstly, the ease of discovering rare diseases through a click-and-upload app coupled with a strong distribution strategy can make the detection of rare diseases simpler and more efficient.
Secondly, the discovery of more patients with rare diseases could help achieve the patient threshold needed to attract the attention of pharma companies. This could make help incentivize drug discovery for the next 10% of rare diseases that have no cure today.
Hopefully, in the future, children like Andrew won’t have to wait for years to get the proper diagnosis and treatment to lead a good quality of life.
Watch my interview with Moti Shniberg and Dr. Karen Gripp of FDNA on how they are innovating with AI to detect rare genetic diseases in children.