[00:00:28] Suchi Saria, Ph.D.: Hi, I’m Suchi Saria. I’m the founder and CEO of Bayesian Health. I also hold an endowed professorship at Hopkins at the intersection of engineering and public health. Today, as a Day Zero advisory council member, I’m extremely excited to interview Dr. Michael Abramoff, a fellow AI entrepreneur. Dr. Abramoff is the founder and executive chairman of Digital Diagnostics. He is also a clinician and professor of ophthalmology at the University of Iowa. Michael welcome. It’s such a pleasure to have you. You’re a clinician and a scientist. You’re also an entrepreneur and you’ve made great strides in bringing a revolutionary tool to market at the intersection of AI and ophthalmology, specifically for diagnosing diabetic retinopathy. I would love to have you tell our listeners a little bit about your background and what inspired this work, and where does it stand in the field today in terms of, is it being used? Who’s using it? And tell us a little bit about the benefits

[00:01:27] Michael D. Abramoff, M.D., Ph.D.: Well, thanks so much for having me, Suchi. It’s great to be here. Very excited to be explaining that. But, let me go back to how this all started for this, what you cal,l exciting new technology. I was in training to become an ophthalmologist. That’s called a residency. And I was seeing many patients with diabetes and you have to know that the most important cause of blindness in the US and many other countries is diabetes, a complication of diabetes. And so these patients typically came in too late, when it’s hard to treat them and to improve their vision. And so we already knew at the time that, if we see these patients earlier, we can do more preventative care. We can prevent this blindness and visual loss. But that was not happening. And I had a background in brain research, mimicking the brain with neural networks going back now 30 years. So I worked on that. I was actually doing a Ph.D. On what we now call artificial intelligence at the time while I was a resident. And so I thought, well, if it can make a computer make this diagnosis, the patients don’t have to wait to come to me, maybe make an appointment, or forget about it. Maybe it’s snowing. There’s other reasons not to come. And instead they can get it where they get the diabetes care. And so, this idea of a computer making a diagnosis seems pretty obvious if you have seen Star Trek or whatever, but it’s costly, in fact, not so obvious to many people that I was going to work with. So that started it and ultimately led to, for the first time ever, FDA clearing or approving an autonomous AI, meaning an AI that makes a decision by itself, no human oversight of this diagnosis for what you already mentioned, diabetic retinopathy. Not only that, but after FDA approval, we spent many years working together with Medicare, with CMS, the Centers for Medicare and Medicaid Services to create a national reimbursement for this autonomous AI diabetic retinopathy exam, which now happens as of January one of this year. You can actually get reimbursed if you use this autonomous AI, which again, there’s no human involved in this diagnosis. So I think, like you said, it’s really exciting that we can now do this because we have shown you can get it all the way from a concept, an algorithm really, all the way to a product for the market fit, but also gets FDA clearance, and then ultimately get to market. And of course we can do this with other diagnoses.

[00:04:02] Suchi Saria, Ph.D.: So this is 2022. Obviously it’s going to be a historic year in terms of you actually being successful in getting reimbursed for this. Walk us through, how long did it take and what were the steps along the way?. So it sounds like the first piece was getting the research to show that you could diagnose these using images. What were the inputs here and were these like hand labeled images from other ophthalmologists doing diagnosis for diabetic retinopathy? How long did it take to develop the system itself? And then it sounds like you did FDA approval. How long did that take? And then you have to go to CMS reimbursement. I’d love to hear a little bit of a timeline.

[00:04:40] Michael D. Abramoff, M.D., Ph.D.: Well, I love to explain this to a fellow AI entrepreneur, as you so nicely said. Thank you for that. And yes, let me take you through what I think is important for those who want to become an entrepreneur or are already entrepreneurs. I was an academician. I was working on, hey, there’s this idea. Let me just prove in a scientific study in a scientific publication that this can be done. And I thought that should be enough to convince every clinician in the world that we should all do this. Well, as you know and as many will find out, that’s not how it works. You need to do many, many publications, and then maybe you start moving the needle. That was still not enough. And ultimately I realized the only way to bring these benefits improving patient outcomes was actually, I need to create a company, effect a quality system, develop the software not in a graduate student level way, but actually in a professional way to bring it to FDA. The first part is really proving that it can be done. And the, in 2010, I had my first meeting with FDA and I came in and I said, well, I want this computer to make a diagnosis. And the FDA rightfully said, wait, wait, wait, let, let let’s talk a bit. Let’s consider this carefully because FDA in the medical field for many years was comfortable with what we now call assistive AI, meaning AI that assist typically a radiologist in making a better, faster diagnosis. This was new because now there would be no radiologist, no ophthalmologist, no clinician involved in the same medical decision. So I think what we worked out together and we later published this, what I now call an ethical framework, meaning, let’s found the ethics of what we do in terms of safety, patient benefits, patient autonomy, what happens to the data, many aspects of AI that people are now talking about and sometimes concerned about there’s sometimes rightfully concerned about. I think we addressed all of them in this ethical framework, but they were discussing and continuing to discuss. Anyway, so working with FDA for eight years on ensuring that we could show safe dealing with AI bias. We did a paper on racial bias in AI, where we tested in Africa in 2013, so long before it became a big thing. We and FDA were really concerned about digital work in every part of the population that you wanted to work on. So with that very collaborative process ultimately leading to the so important market entry in 2018, where the FDA said, after a first-time pre-registered clinical trial for an autonomous AI, where we show it the safety, the efficacy, the lack of racial, ethnic bias and age bias. And so then we were in the market. And that was really new because there was just no computer making a diagnosis. I’m still proud to say that it was the first autonomous AI, period, available to consumers. And there was actually, the Institute for Highway Safety last week published that you cannot buy an autonomous car, that consumers do not have access to autonomous driving today. That is the Institute for Highway Safety. I’m not saying that. So really the only way you can have access to an autonomous AI is through a medical diagnosis where you can walk into a supermarket in Delaware and get your diagnosis from a computer. And so isn’t it interesting that a healthcare system, which is this horribly complex thing, was actually the first to use autonomous AI for essentially better outcomes for patients.

[00:08:18] Suchi Saria, Ph.D.: Walk us through an example of, when this patient walks into the supermarket in Delaware, what are they seeing? What are the inputs? How long does it take? And what is the output?

[00:08:29] Michael D. Abramoff, M.D., Ph.D.: Absolutely. Don’t forget that this is embedded in the healthcare system and it’s solving a very specific narrow problem. Patients with diabetes have their diabetes management. They get their scripts to do the better breath of control, because we all know there’s many other complications from diabetes that you can get and prevents with good treatment. So they have their diabetes management. They do go there. The challenge has always been this eye exam, the retinal exam that requires a referral because the primary care physician, the endocrinologist, is not comfortable or doesn’t have the expertise, or feel they don’t have the expertise to do a good job of these exams. So they refer the patient to an eyecare provider, an O=ophthalmologist, an optometrist, a retina specialist like me. The problem is that it’s not happening. Even today, 15% of all the 30 million people in the US who need an eye exam are getting it. 85% are not. And we know all too well that is eye exam is preventative for vision loss and blindness. So right now, 60,000 people a year are losing vision directly as a cause of diabetes when we can prevent this by just doing more diabetic eye exams. So with that said, if you have your diabetes management, rather than being referred, have to make an appointment, have to travel, it’s snowing, or it’s hot like in Iowa. Or, it’s four hours drive, like in many places, or there’s just no eyecare available in inner city Detroit or New Orleans. Many challenges for accessing these diabetic eye exams. That is now solved because this autonomous AI can literally do that wherever there’s an outlet. So if you’re managing someone’s diabetes as a provider, the healthcare system, you put this in place and you can, within a few minutes, make this diagnosis rather than referring the patient out. And not only that. You can get reimbursed. You close the care gap. So all of that, that’s from the perspective of the provider in the healthcare system. From the patient, it saves you all this travel, having to not forget to appointments, and knowing right there when you’re still with the person who’s managing your diabetes, like a nurse or a physician, they tell you, well, it’s all good with your eyes or, hmm, there’s something concerning here. Now you really need care and treatment. In most patients, that’s not the case.

[00:10:49] Suchi Saria, Ph.D.: Does somebody else prescribe it? Or can you just walk up to a kiosk and self administer if you think it’s appropriate? And if somebody has to prescribe it, is it at a primary care physician’s office?

[00:11:00] Michael D. Abramoff, M.D., Ph.D.: It’s prescribed. It’s a prescription device, very deliberate decision by FDA. As we move AI, and autonomous AI especially, from the lab into clinical practice, let’s do this step by step. So it said we want this to be a prescription device embedded into the healthcare system. We do not want an orphan A,I as I’ve called it, where people just go off on their own and have the diagnosis and someone is paying for it when it doesn’t feed back through the diabetes management. Once you have this complication of diabetes, it’s an important warning sign that your metabolic control needs to be betteer. So it helps the primary care provider, as well as helping the patient to prevent visual loss and blindness from timely treatment.

[00:11:41] Suchi Saria, Ph.D.: Got it. And how long does it take to do the test? Is it pretty quick? Is it something, do they need like hardware for it? That’s actually pretty hard to procure. And as a result, only a small number of centers have it.

[00:11:54] Michael D. Abramoff, M.D., Ph.D.: We we care greatly about what you mentioned, that only a small number of centers have it. So access is key. We want all the 30 million people with diabetes to have this exam. So we typically do it through subscription service. So there’s no upfront costs for the camera or anything. You need a camera, but we made it. So, we actually showed it in the clinical trial, that a high school graduation is enough to use and operate this system. So it’s very simple to use. It takes 10 minutes on average. And so by the time the patient actually talks to the provider managing their diabetes, they will tell you to stop smoking, to eat more vegetables, to take good care of yourself, and also, hey, your eye exam is good or bad, as the case may be. So it’s very efficient in the workflow of the clinic.

[00:12:38] Suchi Saria, Ph.D.: That’s That’s very exciting. So you were describing me your timeline and it sounds like, in 2010, you approached the FDA. 2018, you got FDA cleared. And then, since then, you’ve been to CMS for reimbursement. Walk us through some of the challenges or questions or data you had to show to be able to get reimbursement for this.

[00:12:57] Michael D. Abramoff, M.D., Ph.D.: I think FDA clearance, formerly it’s de novo clearance. We say approval. FDA doesn’t want me to say approval, but de novo clearance is, of course, very important. I think, and we may have joked about this before, my nickname, now 12 years ago, was actually The Retinator, the terminator of the retina, so-called by my colleagues, by the ophthalmology committee, where it was a big editorial about my research saying, oh, this is going to cost us jobs. This is going to lower the quality of care. Should we want this? And so the answer was, no, actually this improves quality of care, improves access, and now the strongest supporters from the autonomous AI for a diabetic eye exam is the American Medical Association, the American Academy of Ophthalmology. So it’s very exciting to see that, if you do it right, if you explain it carefully, if you’re very open, if you show with evidence that it is safe, there’s no bias, et cetera, this can really work and it’s acceptance can be there. So part of the challenge was just essentially, and I mentioned this ethical framework and you could say, well, it’s very nebulous and vague, but it was actually really important to show that there was these ethics that were answered. Now, the bias that we already mentioned, what happens to the data of the patient? Who’s liable, right, we had to solve the liability. If you’re a physician, you’re liable for your medical decisions typically. What about an AI? And so we, early on, said that the AI creator, in this case the company, should be held liable for the performance of the AI. That is now part of the policy of the American Medical Association for autonomous AI. But someone needed to say that because many AI companies try to avoid that issue rightfully or wrongfully. That’s their decision. We decided, if we say this is autonomous, just like a doctor make a medical decision, someone needs to accept liability. You cannot leave that with the patient or the healthcare system, or the provider who is using the AI, who is essentially trying to outsource this diabetic eye exam, right? So that needed to be solved. Many interactions with CMS, the Centers for Medicare and Medicaid Services. They were very concerned, oh, if we allow one AI, will we allow all AIs? Also, the AIs that don’t improve patient outcome that are not ethical, all the concerns with that needed to be addressed. It was very important, I think, that the CPT editorial panel created the CPT code 92229 that is very narrow. It talks about the disease, diabetes. It talks about the diagnosis and the process of the autonomous AI making this decision. So it’s easier to limit it to a very specific CPT code than to say, oh, we’re now going to rebirth all AI. CMS is being very concerned all the time about what they call the guard rails and you can look it up in the federal register. There’s dozens of pages discussing, how do we deal with AI in the future? How can we reimburse it? AI bias comes into it. Does it actually improve patient outcomes? What does this mean for access for patients? Will it increase or improve health disparities? And so all these things need to be addressed for any new AI as well, but at least there’s a path now.

[00:16:16] Suchi Saria, Ph.D.: Absolutely. When you were needing to show access, improvements in access, some of it came from being able to deploy this tool very easily in a lot of different places and the fact that the user doesn’t need to have advanced degree to operate your system. Walk us a little bit through the cost benefit because, very often, when we introduce new technologies, often people view it as increasing healthcare costs because now we just have more expensive tools that are going to increase the cost. Can you tell me from your own perspective as a clinician, do you see this as net decrease in cost? And if so, how?

[00:16:55] Michael D. Abramoff, M.D., Ph.D.: Thank you for asking that. The reason to found Digital Diagnostics was to lower the cost of healthcare, improve quality, and improve access. That’s the reason for its existence. For this case, if you look at the reimbursement, where it’s between $46 and $64 from Medicare, if you compare that to what Medicare pays me as a retina specialist for the same diabetic retinal exam, it’s about $70. So you have an immediate cost saving about two thirds anytime a patient does this rather than the full eye exam. And I like to think I’m a really good retina specialist. But for this specific exam, we only need to know whether or not it’s diabetic retinopathy. And so there’s a cost saving built in for every patient that we do. CMS was rightfully worried, well, if we start doing more of these eye exams, which is benefiting patient outcome, what happens to the budget? And you can see that even if you do, the patients actually require it, it’s still better for the budget. So part of that was the immediate cost savings from a payer perspective. But if you look at the provider, the healthcare system, previously, they had to refer this out. Now they can actually get reimbursed for something. So it’s very attractive from a return on investment perspective to deploy these systems and use these systems. And that’s why we have already implemented in hundreds of sites. We’re going to the thousands of sites in this year and maybe the next year. And it also means we already developed best practices for how to do this workflow because on the one hand, yes, ,it’s easy to use. It’s easy to deploy. It integrates easily into the EHR and the healthcare system. But it offers a new diabetic eye exam, where previously it wasn’t available. What is the best way to not only have patients get a diabetic eye exam because we are a company that is primarily focused on improving outcomes, like I said, and saving costs. So we work with our customers to make sure that the patient outcomes are improved. And we see that in the numbers. Not only do the patients get a diabetic eye exam, they also get timely treatment now if they need it. And in fact, we have studies showing that the patient satisfaction goes up because they don’t need to travel out anymore for the diabetic eye exam. Rather they can stay where they are and they get it immediately. In Parkview, it was a great example of where we’ve shown that in the publications coming out soon. Very exciting to be honest.

[00:19:32] Suchi Saria, Ph.D.: Very very, very cool. So I know at some point, I remember reading that Google was getting excited about pursuing developing a tool for diabetic retinopathy as well. I’m curious, did you collaborate with them? Are there lots of other companies working in this space? One thing that’s really interesting for you guys is you’ve been around awhile, you’ve been working on this for awhile. So, generally as entrepreneurs, you have like a two, three to four year appetite for bringing some new tool to market. Walk us through, from your standpoint, like as big tech entered, was that anxiety generating, was it exciting? What did that do for your investors from a point of view of the enthusiasm for Digital Diagnostics co-op? And then are you seeing a lot of other teams tackling now wanting to attack the same problem?

[00:20:17] Michael D. Abramoff, M.D., Ph.D.: Yeah to start, what you said at the end? Yeah. There’s many teams now doing it, especially because we helped create a reimbursement. Now, there’s suddenly, I won’t say a pot of gold, but there’s definitely interest from an ROI perspective. But going back to the beginning, I was originally a clinician, a scientist, trying to improve outcomes for my patients and other patients. And so I said, well, I will help the patients by having this computer make a diagnosis. I did the science. I founded the compan., I found angel funding. I worked with FDA. It’s very exciting to see that, at the time, the biggest company in the world say,ey, we want to do this too. It’s validates that, hey, there’s something there. So that was exciting to see. It’s exciting. It’s validating. But it’s also, wow, we better really, we’re this tiny company from Iowa, in the middle of fly over country, so to say. And so we really needed to show that we would be winning and first to market. That was essential. And so one big decision we made that I always saw the right way to do is to work within the healthcare system. It’s probably natural because I’m a physician, but I thought, I want to embed us in the healthcare system. And I see the FDA as the highest hurdle to meet in terms of safety and essentially ethics. I want to go through there. The choice was, at the time, why don’t we just try it in less regulated countries, experiment there on millions of patients, and then bring them back here and say, hey, we already did the work. Let’s just go. No, we decided to challenge this. That’s why I started meeting in 2010 with FDA. And again, in the beginning, it was not like, oh yeah, we want this too. Now, it was why did we talk? And so, I liked seeing that our approach working with the healthcare system with patient organization. For example, the standards of care are run by a patient organization called the American Diabetes Association. They updated their standards of care to support the use of autonomous AI. So patient support is… Providers, I already mentioned it, physician organizations, professionals excited, supported. Now regulators supported. Everyone supported it. I think this was really important, for everyone to be able to say, I trust this. I want this to be used on patients. And I think we achieved that. And that’s why it was so exciting to be first to do that. And I think everyone saying we trust this, from patients to physicians, everyone else, also helped CMS make this momentous decision. So I think we did it the right way, but it was a risk and it was a challenge. But it is also very motivating to see that others are following these footsteps now. And not only that, because we talked a lot about diabetic retinopathy, but there’s so many other things we can do with AI and especially autonomous AI in my view. And so there’s now a path. It has been done before.

[00:23:16] Suchi Saria, Ph.D.: Absolutely. And I think it’s very exciting what this next decade is going to look like in terms of number of different areas we can tackle. One thing that’s happening right now is staffing shortages, right, like we’re just seeing nationally, not only shortage of healthcare workers, people are burnt out. Outcomes have plummeted. So this is exactly the sort of scenario where AI-led tools can actually both improve outcomes, but also augment our workforce, even in fully autonomous scenarios or assistive scenarios. I think this makes a ton of sense. I’m gonna switch gears a little bit. To be able to bring research led innovations or research intensive innovations to market, a big part of that is persistence because there’s just so many steps you have to cross, especially when it’s a new market. How did you motivate your team? How did you build your team? Like what learnings did you have from a team building perspective, especially in this current era where people turn over every two years you know, it’s hard to like in an area like this, you need to learn over time. You need to build expertise. Are there any lessons you have for other entrepreneurs, in fact, even lessons where maybe early on, things you learned that you didn’t do right and you had to fix it over time to motivate and build the right kind of team?

[00:24:32] Michael D. Abramoff, M.D., Ph.D.: Yeah, great question. There’s so many lessons learned. There’s no way to answer all of them. I already mentioned, I think, the ethics were very important. The early hires, I interviewed everyone, myself, obviously, because I started the company. And for developers, engineers, you don’t often talk about the ethics. I teach myself, right? So you get this obligatory course in ethics and then, that’s it. And so I really explained to them, we can in the aggregates measure the performance and the safety of this AI. But you will make decisions when you’re coding with a greater than, or greater than or equal sign that you put in your code that may make a difference for one individual patient one day, that is just no way to measure in aggregate because sensitivity is an average, right? There’s just no way that we look at each individual. We talked about it a lot and I think some were surprised by that, but I still hear that people really appreciated that, how important every line of code was. It was not just, oh yeah, you can do it this way or that way. Let’s just make it work. No, it was much more than that. So I think that’s one motivating factor, that we paid so much attention to those aspects, always thinking about what does this mean for patients? Are we actually doing the right thing? That kept coming back. Another thing I learned is that the team you start with may not always be the team that is appropriate for the next step, right? There were a lot of very dedicated developers that are engineers, scientists, Ph.Ds that are very good, still with us, but there was very much a focus on one customer, which was the FDA. We spent almost 10 years saying, whatever the FDA wants, we will do, or what we agree to do together, because if you don’t get through the FDA, there’s no market, there is no other customer, right? And so you needed market access. And so we focused the team very much on that, but that is a product market fit, but the market is only one customer. And so once you get past that, immediately, the market turns out to be obviously, in retrospect, much bigger, very different healthcare systems, providers, reimbursement, care gap closure. All these are the things that in, by themselves, are payers, are customers, that the product also needs to meet. And so you need a team that can address all of that. And I’m very excited to say that, ultimately, two years ago, I was able to expand the team with two really seasoned healthcare professionals, John Bertrand, and Seth Rainford, who came from a very prestigious background. And together, I’ve literally called them co-founders because we really built a team to address all these other things. Again, with developers and engineers, FDA, you can tackle, but after that, the healthcare system is so complex, it’s so varied. There’s so many different constraints and demands on whatever your product is. It’s good to have a lot of expertise there as well.

[00:27:43] Suchi Saria, Ph.D.: I’d love to hear a little bit of your vision for where are things headed next. Like what does the next five years look like for Digital Diagnostics and then perhaps the field as a whole, from your point of view?

[00:27:53] Michael D. Abramoff, M.D., Ph.D.: Absolutely. So for Digital Diagnostics, A, we have so many contracted systems that we need to actually implement. So there’s a little bit of a backlog there. We’re expanding the team as much as we can. And like you said, it’s not easy to find people, high quality people, right now, but we’re finding them and they’re really attracted to come work with us. So we just need to serve our customers. That’s the first thing, make sure not only that we turn them on, but also that we improve patient outcomes, patient satisfaction. So we have these customer success teams literally working with the customer for months and sometimes years to make sure that it all works. And these bigger healthcare systems where they have 30, sometimes a hundred, installs it’s a long-term relationship. So part of that, and then also, there’s so much opportunity now for other autonomous AIs, for other autonomous AI for the eye, for the retina, for other organ systems like the ear, for skin, we already have one there. We’re going through to start an FDA trial very soon now. So we know how to do the trial. We know how to get the FDA approval, de novo clearance. We have done reimbursement before, so we’ll renew again and again, and again. We will also see other companies, we’re already seeing that, now that it reached reimbursement, now that it’s been shown that you can get FDA de novo clearance approval, now other companies are of course following, including the biggest company in the world at the time that we just discussed. So I think a lot is going to change over the next few years, a massive expansion of how many people are getting the diabetic eye exam finally. That’s the biggest thing. And doing, repeating that for other AIs, for other disease categories. It’s just a very, very exciting time, to be honest.

[00:29:41] Suchi Saria, Ph.D.: It is indeed. Michael. Thanks so much for making time, for sharing your very deep and wide experience in this space. I’m so excited to see Digital Diagnostics expand, but also what I see in other things shifting in this space is in the last decade, early on the way AI entrepreneurs were mostly technologists, but they didn’t really understand healthcare. And I think a big shift we’re seeing now are individuals like yourself, research led scientists, deep understanding of the healthcare system, deep understanding of healthcare data sets, and AI tools that really embrace the complexity of medicine. And I think that’s absolutely vital for succeeding in this space. So thank you so much.

[00:30:22] Michael D. Abramoff, M.D., Ph.D.: Thanks so much, Suchi, it was great.