Generative AI and Rare Disease
About a quest for a cure, face recognition tech, and a failed project.
This is the unusual story of The Yellow Brick Road project, a project that has failed.
And even though it failed, we have learned so much from this project.
Many of our products started as hackathon projects. In former blog posts, I wrote about a couple of successful ones, like Project New Hope. But sometimes projects are not successful. This is the story behind one of those.
The Yellow Brick Road
It was July 2019, shortly prior to the company annual hackathon, when my colleagues Yael Gruenbaum-Cohen and Ofir Tessel introduced me to a non-profit organization named The Yellow Brick Road.
Doing a project for a non-profit organization is something we sometimes do in hackathons. We typically end up donating the code and all the by-products of the project to the non-profit. And that specific project supported a great cause.
The team and I were introduced to the parents of a little girl named Yaeli.
Yaeli was born with a rare genetic disorder caused by a mutation in the HNRNPH2 gene. The mutation causes developmental delay, intellectual disability, muscle tone abnormalities, autism, and seizures. The disorder exclusively impacts young girls since the genetic mutation is associated with the X chromosome.
The mutation disrupts the entrance of protein into cell nucleuses, preventing the delivery of vital proteins that are critical for child development.
Multiple treatment approaches could be researched, but the HNRNPH2 mutation was an Orphan disease: there were just not enough patients in the world to make it commercially worthwhile for pharma companies to develop a drug for the condition.
The mutation was newly discovered in 2016. At the time we started the hackathon project, only 78 girls were found to carry this mutation. The only way to identify HNRHPH2 was via whole exome sequencing (WES). In most countries this test wasn’t subsidized, and many families couldn’t afford the test. To kickstart development of a cure, pharma companies needed at least 300 identified cases of the mutation.
The Yellow Brick Road Project was a not-for-profit organization, that was founded by parents of girls affected by the mutation, to promote finding a cure for this rare genetic developmental disease.
The team and I immediately fell in love with Yaeli and her family. Yaeli was a sweet little girl with freckles and a lovely smile. She had a beautiful, supportive family, and it was obvious they were dedicated to Yaeli and loved her with all their hearts.
The Hackathon Project
It so happened that all the existing HNRNPH2 patients had similar facial features. The girls had special shaped eyes and round eyebrows, resulting in physical facial resemblance.
The idea behind the hackathon project that the non-for-profit asked us to build for them was to try to leverage Face Recognition AI to score faces for similarity. Our goal was to help find more patients, to incentivize pharma companies and make it commercially viable for them to develop a cure.
And so, we kicked off the hackathon project, that borrowed its name from the name of the non-profit organization, and worked with The Yellow Brick Road to develop a tool that would help identify more patients and flag them to get their exome sequenced.
In that hackathon we have built a service that scans faces for those unique features and scores faces for similarity, to help find more patients around the world. We have also built a patient engagement experience that aimed to help the not-for-profit expose the tool, raise awareness and connect with other patients.
What went wrong then?
In hindsight, there were so many red flags, that I’m asking myself what on earth were we thinking. Here are some of the issues.
First and foremost, we did not have enough data. Rare diseases, by definition, have a limited number of patients, resulting in scarce data for training AI models.
It is crucial to understand that the success of such face recognition project depends on the quality, quantity, and diversity of the data used to train AI models. We now know that we needed much more data, of real patients, but just we did not have enough of that.
Diverse patient images:
We needed larger, more diverse set of facial images of patients with the HNRNPH2 mutation. This includes images under various conditions (lighting, angles, expressions) to train the model to recognize the key features.
We needed a significant dataset of images of individuals without the mutation, to serve as control group, ideally matched for age, ethnicity, and gender, to teach the AI the variance in normal facial features and reduce false positives.
We needed to include images of individuals with other genetic mutations affecting facial features, such as Down Syndrome, etc. That would have helped the AI learn to distinguish between HNRNPH2 and other conditions, enhancing specificity.
Comprehensive patient data:
We needed patient data that would include detailed genetic information: beyond just identifying the presence of the HNRNPH2 mutation, detailed genetic data and potential interactions with other factors could enhance the model accuracy and predictive power. Comprehensive clinical profiles, including symptoms, progression of the disease, and response to any treatments, would have been critical, too.
Ethnically and geographically diverse data:
Given the global scope of the search for patients, incorporating facial images from various ethnic and geographical backgrounds would have been crucial. This diversity could ensure the AI effectiveness across populations, considering the variability in facial features influenced by ethnic backgrounds.
Longitudinal data:
Capturing changes over time through longitudinal data on patients with the HNRNPH2 mutation could have offered insights into the progression of facial features and other symptoms. This could allow for more dynamic models that can account for age-related changes or the progression of the disease.
Anonymized data sets:
Given the sensitivity and privacy concerns with genetic and facial recognition data, especially involving minors, anonymized datasets for all of the above, that protect the identity of the participants while allowing for meaningful analysis and training of AI models, would have been essential.
These types of data would have been critical for improving model accuracy and reliability, reducing bias, enhancing Specificity and Sensitivity, and adhering to Responsible AI standards.
Ethical considerations:
The use of face recognition technology, especially involving children, raises significant ethical questions. Concerns include privacy, consent, and the potential for misuse of the data. Given our high Responsible AI standards as a company, this would have been a huge barrier.
Regulatory considerations:
One more issue that served as the last nail in the coffin for this project was that the AI model was looking more and more like a diagnostic tool, making it too close to become a Medical Device. Software as a Medical Device (SaMD) is defined as software intended to be used for medical purposes and performs these purposes without being part of a hardware medical device. This can include software services or applications that are intended to treat, diagnose, cure, mitigate, or prevent disease. The classification of such an AI tool as a medical device would introduce a layer of regulatory complexity. Such tools must undergo rigorous validation and approval processes before they can be used clinically, which would have been a barrier to quick deployment.
The project got buried, and never saw the light of day.
But one of the by-products of the project was a video clip we produced for the project. The company donated the production costs, the video was shared publicly, and we were later told by Yaeli’s parents that the video helped them raise awareness and support for the cause. So we were able to help a little.
The Vow
Daruma dolls are Japanese traditional dolls that serve as a symbol of perseverance and good luck. When you buy them, the doll's eyes are both blank. You select a goal or a promise, and you paint the left eye of the doll. The doll then sits on your desk looking at you with that one eye, constantly reminding you of your promise. Once the promise is achieved, the right eye is painted.
I got such a daruma doll when I founded my organization back in 2015. The vow I made to my doll when I painted its first eye was that I will do everything I can to help patients and make a difference in people’s lives, even if it meant helping just one patient.
The doll was sitting one-eyed on my desk at the office for 5 long years, and I never felt I did enough to earn its second eye. But when Yaeli and her family visited our team at the office, Yaeli gave my doll its second eye.
Could Generative AI assist with the gaps?
Recent tech evolution raises the question on whether Generative AI could have made things easier for this project. For example, could Generative AI be used for generating synthetic images, to overcome the data gaps?
Intuitively, it sounds all wrong to use Generative AI to generate images for training a face recognition tool that is supposed to identify a rare genetic disorder - it’s like you’re asking the model to teach itself. Let’s spell out some of the key problems explicitly:
Generated images suffer from lack of authenticity. They may not capture the real and subtle variations in facial features of real patients, reducing the model's ability to reflect real-world conditions.
There’s a risk of introducing - or amplifying - biases that are related to various ethnicities or to the manifestation of genetic disorders in patients. In other words, Generative AI would generate images based on stereotypes, thereby reinforcing and perpetuating its own biases.
The generated images might not accurately reflect the diversity needed for a model that is supposed to work globally, in terms of coverage of ethnicities, mutations and more.
Simulating genetic interactions is complex - Generative AI would have a hard time synthesizing multiple genetic mutations and the clinical manifestations associated with the mutations, as potentially reflected in facial features.
The conclusion is inevitable, all the above challenges that caused the project to fail would still be there, and then some more.
How it all ended
The project has failed. We were not able to proceed with it, and were unsuccessful in helping Yaeli. But she touched our hearts, and we still think of her often.
The parents from The Yellow Brick Road Project taught us that the battles of life are not won by the strongest or fastest, but by those who never give up. They inspired us all with their perseverance and love for their daughter.
Yaeli passed away last month, February 2024, after suffering complications that were related to her genetic condition.
She will always remain in our hearts.
In the loving memory of Yaeli Farkash, 2013-2024, may she rest in peace.
About Verge of Singularity.
About me: Real person. Opinions are my own. Blog posts are not generated by AI.
See more here.
LinkedIn: https://www.linkedin.com/in/hadas-bitran/
X: @hadasbitran
Instagram: @hadasbitran
Recent posts:
I really like this post, not only because it's touching and inspiring to read about your efforts to help young girls inflicted with this rare disease. But also for sharing your lessons learned with such clarity on why it failed. These learnings are important for those of us working on AI-powered diagnostics. The challenges are real around sufficiently representative data, especially for less prevalent diseases. And yet, we must persist. Thank you!