Dekel Gelbman and How Machine Learning is Changing Rare Disease Diagnosis
EPISODE SUMMARY
Harry’s guest is Dekel Gelbman, founding CEO of FDNA. The company uses a combination of computer vision, deep learning, and other artificial intelligence techniques to improve and accelerate diagnostics and therapeutics for children with rare diseases.
Please rate and review The Harry Glorikian Show on Apple Podcasts! Here’s how to do that from an iPhone, iPad, or iPod touch:
1. Open the Podcasts app on your iPhone, iPad, or Mac.
2. Navigate to The Harry Glorikian Show podcast. You can find it by searching for it or selecting it from your library. Just note that you’ll have to go to the series page which shows all the episodes, not just the page for a single episode.
3. Scroll down to find the subhead titled “Ratings & Reviews.”
4. Under one of the highlighted reviews, select “Write a Review.”
5. Next, select a star rating at the top — you have the option of choosing between one and five stars.
6. Using the text box at the top, write a title for your review. Then, in the lower text box, write your review. Your review can be up to 300 words long.
7. Once you’ve finished, select “Send” or “Save” in the top-right corner.
8. If you’ve never left a podcast review before, enter a nickname. Your nickname will be displayed next to any reviews you leave from here on out.
9. After selecting a nickname, tap OK. Your review may not be immediately visible.
That’s it! Thanks so much.
Transcript
Dekel Gelbman and Using AI to Reduce Time-to-Diagnosis
Note: MoneyBall Medicine is produced for the ear and designed to be heard. If you are able, we strongly encourage you to listen to the audio, which includes emotion and emphasis that’s not on the page. Transcripts are generated using a combination of speech recognition software and human transcribers and may contain errors. Please check the corresponding audio before quoting in print.
Harry Glorikian: Welcome to the Moneyball medicine podcast…
I’m your host Harry Glorikian. This series is all about the data-driven transformation of the healthcare and life sciences landscape. Each episode we dive deep through one-on-one interviews with leaders in the new cost-conscious value-based healthcare economy. We look at the challenges and opportunities they’re facing and their predictions for the years to come.
My guest for today is Dekel Gelbman, who is the founding CEO of FDNA. He leads the corporate and business strategy of an innovative digital health company that develops technologies and SAS platforms used by thousands of clinician’s researchers and lab sites locally in the clinical genomic space. The main mission of the company is to give hope to children with rare diseases and their families.
FDNA which was founded in 2011, uses a combination of computer vision, deep learning, and artificial intelligence to analyze patient symptoms, physical features and genomic data in combination with a network of thousands of genetics professionals worldwide. Then they drive scientific insights to improve and accelerate diagnostics and therapeutics impacting the lives of children with rare diseases.
Harry Glorikian: Dekel, welcome to the show, good to have you.
Dekel Gelbman: Thank you very much, it’s a pleasure being here.
Harry Glorikian: Dekel, we’ve known each other almost since the day you showed up here in Boston deciding whether you would place yourselves here as a company. Tell me how this whole thing got started, because it’s not exactly what you would consider a normal route into the world of diagnostics or using AI and machine learning, and it was quite a while back. I mean it will you were guys were at the forefront of this before I think a lot of other people got involved.
Dekel Gelbman: Absolutely you know, when we started we knew almost nothing about healthcare. We were techies, the background of this company was actually two founders that were very successful in developing facial recognition software that was sold to Facebook in early 2010. And the drive, I think for this company was how do we make an impact, real social impact with this technology or with our know-how around facial recognition. And so by exploring a lot of fields, Healthcare was really very compelling.
Because of the impact that you can, you can make and we started to meet with various specialists and different practices in health care. And then almost by accident, we stumbled across genetics and we were amazed to learn that back then and for decades’ geneticists would look at faces of patients and make a lot of the diagnostic choices based on facial patterns that they could identify. And it was just a lightbulb moment right, then there we understood that we can really drive change, we can disrupt this entire field, we can really drive with a strong computational basis diagnostics. And that was really the genesis of FDA how we started.
Harry Glorikian: Yeah, I remember when you guys we were sitting at what was it Henrietta’s Table at the Charles Hotel and I said you guys told me this and I was like, oh my god that’s just brilliant. I was like, and I always thought it would be direct to the patient. But you guys decided to go to the clinician and come about it from a sort of group learning, group educational perspective on how you teach the system. Tell me a little bit about how it’s designed or and how its deployed and how it keeps learning?
Dekel Gelbman: So with AI, I think today even more than ever it is very obvious that it’s a data play. The more data you have the better the data is the better the technology can become. Learning algorithms and especially today with deep learning models, if you have enough data and the data is good, you can train a very accurate and advanced technology. But the problem in the challenges in this world, especially with rare diseases and genetic disorders is access to that data, how do you get data. When we started, we started with a lot of collaborations with different researchers around the world and everyone was very enthusiastic, but every single research site had only very limited quantities of data.
And so it got us thinking you know what’s the best way to start gathering all the data – collecting, curating it. And I remember, it was one of our developers who said you know everyone uses iPhones right now, let’s develop an app and ask all the geneticists around the world to help us annotate data and collect data. And we said you know, let’s give it a try and that’s how face2gene our current platform was born, and in hindsight you know several years after launching face2gene, this was a very successful strategy.
We were able to deliver an application that produces real-time value clinical value to clinicians and in return and we distributed it for free by the way. In return, we got a lot of data, and we were able to really advance our development of the technology significantly, because of this strategy.
Harry Glorikian: Well, I need interestingly enough if I remember our conversations correctly, it wasn’t just the acquisition of data but it was having experts in the field constantly teaching the system how to be more accurate by their experience.
Dekel Gelbman: That’s the old AI. So, when we started really supervised learning or having experts teach the system, how to think was how we started, how people thought about AI at the time. In 2014, there was a different trend towards deep learning, where you really don’t teach the system anything the computer identifies patterns on its own. It’s sort of a black box and that’s some of the criticism towards AI today is that being a black box. And that made curating quality data even more important more significant to that process because we no longer influence the system’s method of learning.
So, everything that we influence is, how we collect the data, how we ensure the quality of the data and how we feed the system with data to avoid biases, overfitting, and a lot of the different problems that AI presents today with deep learning.
Harry Glorikian: Can you give me some examples of where this has really changed a timeline, improved that diagnostic Odyssey? How that’s affected you know a patient or a family, and where do you see this, you know where do you see going from a cost perspective and so forth?
Dekel Gelbman: Absolutely, so you know it’s very hard to give macro examples or macro data about time to diagnosis, but on a case-by-case basis we hear all the time from our physicians, from physicians using face2gene, how this integrated into their workflow? How it simplified the workflow? How it helped them choose the right diagnostic tests? How it helped identify specific diagnoses for patients that were looking for a diagnosis for years? So, there are multiple examples and they’ve been published elsewhere both in scientific publications and the media.
But I want to tell you is what we’ve learned in our journey, because when we, you know as you articulated that in the beginning, the mandate that we had going into this journey was how can we help physicians identify or diagnose rare diseases in pediatric settings earlier. And as we started to gain traction as more and more hospitals started to use this as part of their workflow, as more and more researchers started to use this technology to make discoveries. We started hearing back from the laboratories, and this coincides with more accelerated adoption of next-generation sequencing.
The labs are starting to offer exome sequencing and whole genome sequencing to physicians as the primary genetic test. But they came back to us and said, listen we get too much information we generate too much information when we do an exome sequencing. And so we want clinicians to really adopt this as a test because of the broad coverage, we need to make sure that when we analyze the results we present to them results that are relevant, clinically relevant. And so it’s not reasonable to present to a clinician, a thousand different variants that may or may not be pathogenic meaning that they may cause a disease or not.
We need to be able to present with that to them a short list of variants that may be causing a disease. In order to do that, we need to integrate what we call our jargon, calls phenotypic information, phenotypic being the information that captures the clinical observation of a patient. Is the patient tall, does the patient have certain clinical symptoms and does the face present certain patterns that are linked or associated with these diseases? And guess what face2gene captured a lot of this phenotypic information as part of the clinical visit, the clinical evaluation. And then it dawned on us that you know we really hit something.
We started to investigate this further and we’ve participated in the study called PEDIA, that aimed to prioritize exome sequencing results based on facial analysis. The results were staggering we showed that for this cohort of patients, for this group of patients that had monogenic disorders that manifest in facial analysis. We can improve the diagnostic rate from about 40 percent to almost a hundred percent, and at that point, the term next-generation phenotyping was born and adopted by us as where we’re going with this company.
We realized that if we offer a computer-based way, an AI best method to look at a patient and correlate that with the patient’s genome, we would be able to pinpoint with very high accuracy, the disease-causing variants. And you’re talking about cost, you can imagine what this does to this entire industry or the potential of what this can to the entire industry. This can facilitate genome sequencing for the entire population, and it now makes sense because we have a scalable approach into how to analyze and interpret genome sequencing data for the entire population.
And this could have a lot of impact on the future of precision health or precision medicine and that is obviously going to have a huge impact on cost. It’s very hard to predict right now what that impact is going to be, and obviously, if we are to pursue this path, we need to go well beyond just a facial analysis, we need to look at the holistic phenotype of a patient. So, that’s where we are right now and that’s the journey ahead of us.
Harry Glorikian: So, when you were building this, tell me some of the experiences or lessons that you learned. You know you originally said, you know we were working on algorithms then we went to a black box machine learning system and you’ve worked it into the physician’s workflow. Give me some of your experiences on what it really took to get this to where it is today.
Dekel Gelbman: I think you touched on that, the last point I think is the most important one and the most difficult one in healthcare today is integrating with workflow. It is almost unimaginable to change the workflow of a caregiver. They’re just too darn busy and trying to, re-educate them is never going to work. A lot of startups are trying to circumvent the healthcare provider. We don’t believe in that future; we don’t think that providers would disappear. We just think that their role is gonna change and so our strategy was how do we empower the caregivers; how do we empower physicians. And we do that by giving them pertinent data and giving them the ability to make educated decisions.
So, we’re helping physicians and they’re grateful and the community of clinical geneticist or medical geneticists really embraced us. Because we were giving them something that they were missing for years and years, and so we actually saved him a lot of time. The traction and the responses and the endorsement that we received from the physician was where we were focused, I would say in the last four years, really how do we give, how do we provide tools that are useful. And you know a lot of this is exploration, we develop something, we test it, we get feedback from the clinicians sometimes they love what we do, sometimes they don’t. But they’re very open and they’re very responsive.
So, for us, that is probably one of the biggest assets that we have as a company is our relationships with our user base. And that really was important in our approach of, how we develop this technology. Everything is driven by what can be useful for our target audience. We learned along the way a lot of things and there are a lot of challenges. Workflow was one, right so how do we give the physicians the flexibility to use these tools and technologies without changing their workflow. Privacy is a huge issue and physicians are probably the gatekeepers for a lot of the privacy regulation around the world.
I’m talking about HIPAA and today GDPR are. The patient privacy is very important and it looks as though the last gatekeeper is the physician and they’re doing a tremendous job. But we had to step up and improve our entire process. And go through compliance processes and ISO certification. Today we’re ranked one of the highest ranking scores on AWS as in terms of our security and privacy infrastructure, but it took a lot of effort. Another thing that we’ve learned I think is how to be ethical in AI. And this is a I think a hot button today specifically in genetics, along the years most of the data that was curated was curated for Caucasian populations, and this created a huge gap in our knowledge our medical knowledge as a society on other ethnicities.
And so we made it a point to diversify our database so that we can be used not only for the Caucasian population but for ethnicities in Africa and Latin America and the Asia Pacific. And this made a huge difference by the way, not only did it made us grow our presence and today were being used in over a hundred and thirty countries around the world but it actually improved our AI. And this is a very interesting thing that I’ve learned along the years. When you train the system to look at different ethnicities, the morphology the way the face looks can be influenced by a variety of influencers. The ethnicity obviously environment can change how your face looks, not as much with the pediatric population but still and your genetics influence how your face looks like.
So, you have to discount some of these factors and by training the system on a very diverse ethnic population, you’re basically taking off the table the differences that relate to ethnic origin, and you focus on the pathogenic morphology, only the morphology, only these patterns that are caused by those genetic disorders. So, just account a few things that we’ve learned along the way.
Harry Glorikian: How big of a data set do you need to or where are you guys now, compared to where you know it was just a few years ago? I imagine that acquiring this data because of the app is much easier, the amount of data that you’re able to get in is significantly higher than going out there and trying to do this yourself or coming up with a specific piece of instrumentation necessarily to do this. And then it was just recently that you guys started incorporating the genomics part of it, and the announcement was not that long ago. But, how do you see that working into the success of the company?
We what we always try to come up with some special piece of technology whereas I feel like the tech world is moving so fast forward, and what it’s bringing is pretty damn good quality and it keeps improving thinking of you know the iWatch and the detail you can get off of an iPhone camera and so forth. So, how do you see that playing a role in what you guys are doing?
Dekel Gelbman: So, you know again one of the challenges at the outset of the company was dealing with very small amounts of data. Our target number of diseases just with the facial analysis technology is somewhere between 2,500 and 4,000. And for each of these diseases sometimes there are only five reported cases in the history of publications. So, we’re working with extremely small sets of data, for us that was a technology challenge that we’ve addressed through some methods like translational learning, where we learn from bigger data sets. And then we take that back to a smaller data set and apply what we’ve learned but generally speaking we work with very small data sets across or for each specific indication.
Face2gene was very successful in gathering more and more information to date, we have more than 120,000 patients that were processed and analyzed through face2gene obviously that enriches our database. The pace of uploading more and more patients into the system is increasing every month, and so I wouldn’t be surprised if in two to three years we will actually reach around a million patients processed through this system. So, that really enhances our ability not only to improve the AI around identification of specific phenotypes but also broadens the coverage, so we can see more and more diseases.
And you were talking a little bit about other sensors like the iWatch. Part of our next-generation phenotype in approach is indeed to enhance our collection from beyond just a facial data into other phenotypic data. So, vital signs that are collected through wearables are part of that, video processing even voice processing. So, the voice can be a very strong indication for certain diseases. Obviously, medical device information that is collected through existing medical devices and medical imaging, all this information should be funneled into a central location that will be able to improve our insights.
Now there are a lot of companies out there that are doing similar or have similar efforts. Our unique approach is that we take all this information and the sole purpose of that is to then look at the genome and try to identify the disease-causing variants. We’re not developing radiology decision support tools or not developing agent diagnostic devices. Our sole purpose is to look at this information say, how can from this information we would be able to infer insights from the person’s genome.
Harry Glorikian: So, you had started this with you know we’re a bunch of technology guys that sort of stumbled into the world of healthcare. What are the experiences you can share as, you know what type of people do you need on the back end doing the coding, doing the work but then integrating that would say people who might be knowledgeable in the disease state and sort of making that whole thing happen? And you’re not all in one place, you have different sites and so that whole process is there of lessons you can share or the magic you can share to help bridge that gap.
Because I always feel that technologists can code, but you need somebody that understands that health dynamic, that disease state, that workflow and then to have to somehow almost meld into one person to be able to produce something that is usable.
Dekel Gelbman: I wish I had a formula, it’s not very easy to quantify what you need in order to succeed. I would say that generally and this is something that I truly believe in, disruption never comes from within an industry. It takes an outsider to look at something and try to solve a problem that exists for many years. At the same time, without the relationships that we’ve created over the years and without the involvement of medical geneticists in our company, we would have never understood the breadth and the depth of the problems that we’re trying to solve. So, for us the AI approach was very straightforward, but going into diving into the details. it started to become extremely complex in terms of how the syndromes are categorized, how genetics works and that’s information that we simply didn’t have.
But as we dove deeper and deeper with the support of many experts in the genetics field and we have an extremely broad and involved scientific advisory board. If you take a look at our website, it’s probably about 30 to 40 people that are involved, we don’t pay them. They’re there volunteering because they really believe in the future of this technology holds. Without their involvement we would have never succeeded to put technology to solve a problem. And without naming names, you know there are other companies out there that are very sophisticated and considered very prominent in the machine learning world.
I think their approach to involving the industry is wrong, taking just one or two sites to train a system or two to be the domain expert is not the right approach. You have to broaden the scope as much as possible, that’s what we’ve done. We’ve been working with almost everyone in this field.
Harry Glorikian: Well yeah I mean, I think technology lends itself to or the technologies these days lend themselves to. I don’t want to say crowdsourcing but you can get a much larger set of input if you’re managing this correctly. When you’re hiring people or when you’re looking at certain skill sets that weren’t on board, how do you think about that. Where might be some of the places that you’d look to find these individuals aren’t falling off trees and if you were in the Bay Area, you’d be fighting tooth and nail for you know the person that hasn’t even graduated yet. So, how are you taking on the right people and finding the right skill sets?
Dekel Gelbman: So, you know especially in the algorithmic development world, talent is extremely expensive, whether it’s in the Bay Area, whether it’s in New England or whether it’s in Israel. These people are extremely expensive, the competition over recruitment is fierce and we’re competing with some you know 800-pound gorillas in the market Amazon, Facebook, Google etc. The one thing that we have in our company that I’ve rarely seen in other companies is a purpose. And so this is a highly marketable trait for a company when you’re recruiting, getting people on board that believe in the purpose of the company, believe that they can make an impact.
I think is such a powerful thing to have as a company, and coincidentally that’s the kind of trait that I’m looking for when I hire people. So, the experience is important and dedication, diligence, intelligence all these traits are very important. The number one trait for me though is passion because I truly believe that if you’re passionate about what you do and if you enjoy what you do and if you believe in what you do, then you’re gonna put you know more from yourself into the company. You’re gonna be more productive, you’re gonna care.
And so that is probably the number one trait that I’m looking for when I’m hiring people, and that doesn’t have to do with geography or with where you went to school. It’s just you know it’s what you care about, and so it’s not that rare to find employees and talent that connect to the mandate of the company that believed in our vision, and recruitment has never been a huge issue for us.
Harry Glorikian: So, where do you see the company going next, from a technology evolution perspective, from clinical impact perspective and then you know sort of your vision beyond that. But sort of those two things I think the incorporation of technology these days is almost like a race, where you’re constantly trying to keep up with the next chipset that’s incorporated, the next software improvement that’s coming faster than I’ve ever seen it in any other time. And then clinically, where do you see that going?
Dekel Gelbman: So, I think we have to be modest in our perspective on the impact that we can make and we need to be cognizant of macro-economic changes in healthcare that we have very little influence on. So, we need to look from the sidelines and try to evaluate where this field is going. We are strong believers that, we are entering into an era of precision health, we’re strong believers that the main driver for that is genomics. We obviously believe that AI is a driver for these huge data sets and what we can do with them. And so within or from that insight, we believe that if we focus but really focus very hard on developing the best technology that regardless of time. I know that’s a huge issue for startups right, but regardless of time whatever, it takes one year two years or five years. We need to focus on making this technology a standard of care alongside genomics and doing that for us means, focusing on value, showing value demonstrating value, showing how we can improve the benefits for all the stakeholders involved in our little space, which are physicians, researchers, labs obviously patients and then life science companies.
Harry Glorikian: If I read that correctly you’re looking beyond the rare disease space.
Dekel Gelbman: I think the immediate value of what we’re doing right now applies to the rare disease space. But the future implies that genomics is gonna play a key role in risk assessment for more complex and also more common diseases. As we start rooting ourselves into the genomics field, yes we see ourselves tagging along to that journey and going beyond rare diseases in the future into almost all diseases. But there’s a huge gap that genomics needs to catch up to apply to other diseases.
Today I think you know mostly genomics is applied to rare diseases, oncology and that’s pretty much where most of the genomics is focused right now.
Harry Glorikian: Yeah, I’ve always thought about some of the stuff that you guys are doing and saying well what if we just started applying that to a broader population. You know we call it a rare disease it seems to manifest itself in, what might be categorized as an issue or a problem or how it hinders someone from you know the life that they want to lead etc. But I want to say that there’s, the deviation of that is you know, there’s probably people that you call normal that probably have some of these traits that we’re just they’re subtle. So, you don’t pick up on them.
So, I always wondered at the application of technology to the broader population.
Dekel Gelbman: I would argue that naming rare diseases is a huge disservice to these type of diseases. If you think about this if you think about the future of precision healthcare every disease is rare, because every disease is gonna be categorized as a unique subset of interactions between different biological systems and mechanisms. And so I think that in 20 years the term rare disease is gonna be obsolete because we will look at every single disease as a unique set of genotype-phenotype and other biological input or feeds into a computerized system, that’s gonna analyze everything.
So, yes today we focus on rare diseases, we focus on the genomic side and, but that’s I think that’s gonna change along the years. We definitely look at FDNA on a very long term scale, we’ve always been able to do that with the support of our investors and the founders and even our employees. And I think that this is the right way to look at a startup.
Harry Glorikian: Anything I haven’t asked you, words of wisdom you know experiences that you want to share before we sign off?
Dekel Gelbman: I think you’ve done a great job, thank you. It’s always a pleasure to talk to you Harry and hear your insights on the world of health care and how that’s developing. I think, we have the privilege to be operating in a very unique era. And hopefully we’re gonna benefit from good timing and we’re gonna seize the opportunity as a company. But even more important than that I really hope that the effort that we’re doing with developing this technology is going to create a huge impact on patients.
Harry Glorikian: Yeah, I do believe in it’s interesting, yeah I’m not sure that the algorithms are the secret sauce or the machine learning back-end or so forth. I feel like some of that is always going to be able to be reproduced by someone else. But the data set I believe is gonna have tremendous value and the impact that it has going forward. So, on that note I want to thank you very much for joining today, and look forward to continued dialogue and updates in the future.
Dekel Gelbman: Thank you very much very, Harry.
Harry Glorikian: Take care, and that’s it for this episode.