The Ophthalmic Project: Dr Abramoff Discusses Artificial Intelligence in Eye Care
In this interview, Dr. Michael Abramoff discusses the impact and future potential of artificial intelligence in ophthalmology. He highlights AI's ability to enhance productivity, improve patient outcomes, and address healthcare access issues, particularly in diagnosing conditions like diabetic retinopathy. Dr. Abramoff emphasizes the need for ethical frameworks, collaboration among stakeholders, and a balance between technology and traditional clinical practices to ensure AI's successful integration into eye care.
Mark Dlugoss:
Artificial intelligence is one of the hottest topics in ophthalmology and vision care today. Its potential is expected to revolutionize how eye care clinicians diagnose and treat their patients. The question is has AI reached its peak in its technological development stage and is it ready for real world application in all segments of ophthalmology?
Hello, this is Mark Dlugoss, senior contributing editor for Ophthalmology 360. Welcome to The Ophthalmic Project, powered by Ophthalmology 360. In today’s Ophthalmic Project, we visit with a leading expert in artificial intelligence who can answer all the questions regarding this incredible technology and explain how AI is and will change the direction of ophthalmology.
Joining The Ophthalmic Project is Michael Abramoff, MD and PhD. Dr. Abramoff is the Robert C. Watzke MD Professor of ophthalmology and visual sciences at the University of Iowa. He also maintains two other joint appointments at the university, as professor of electrical and computer engineering and professor of biomedical engineering. Finally, Dr. Abramoff is the founder and the executive chairman of Digital Diagnostics, an AI diagnostic company that was the first in the medical field to receive the FDA clearance for an autonomous AI system that instantly diagnoses diabetic retinopathy and diabetic macular edema at the point of care.
Dr. Abramoff, welcome to The Ophthalmic Project.
Michael Abramoff, MD:
Hey, thanks so much for having me, Mark. Very exciting.
Mark Dlugoss:
I’d like to start our discussion with a brief primer about artificial intelligence. Basically, its history, principles, the goals in which trying to utilize it in medicine, specifically in ophthalmology maybe.
Michael Abramoff, MD:
Absolutely. What I like to talk about when I mention AI, artificial intelligence, is that it does what I call cognitively complex tasks that were previously only done by humans, such as ophthalmologists, such as retinal specialists that are experts, and are now done primarily or only by a computer.
Mark Dlugoss:
What are the advantages and disadvantages of artificial intelligence in general?
Michael Abramoff, MD:
I think you need to get it right. The first thing is how do you know whether you’re doing it right or wrong? We’ll get to that. Then more importantly, you need to look at what the problems are in healthcare, and in eye care specifically. In my view, those are productivity, loss of productivity, and lack of productivity, which leads to rising prices and lack of access, and therefore impaired health equity. It also has to do with efficiency and quality of care.
We can improve quality of care, we can improve productivity, we can improve clinical outcomes, we can improve patient satisfaction if we get this AI right. But if we do it wrong, we will harm patients. There is evidence that, not in ophthalmology, in other fields, that has happened already. Again, getting it right is key. Gaining trust and retaining trust is key.
Mark Dlugoss:
Obviously, one of the things going on not only in the medical field, definitely in ophthalmology, is the shortage of physicians and ophthalmologists. How does AI appeal for this advancement when it addresses this problem? What other social economic demographics that could contribute to the growth of artificial intelligence?
Michael Abramoff, MD:
One of the biggest problems in ophthalmology right now is access. There’s many numbers out there, but in the case of diabetes, but in many other diseases including glaucoma as well, most patients who need our care do not even get to us. That’s the access problem. We think about all the patients we see, as clinicians. We don’t think about all the patients who never get to us in the first place. Key for AI, in my view, is bringing those patients in, getting those patients to get high quality care. That is, I think, the first thing.
Then you mentioned a lack of physicians, a lack of ophthalmologists, or lack of retinal specialists, my specialty, and it’s very exciting that very recently, we published the results of a randomized clinical trial comparing AI to no AI, autonomous AI, the product that we created, in Bangladesh in retinal clinics. It showed, to the productivity of those retinal specialists, thanks to the AI, improved by 40%. That may sound a lot to you or maybe not so much, but it’s the same productivity gain as it’s taken 50 years in any other industry in the US to have that same gain. But you can do it in one fell swoop with AI, in one of the lower income countries in the world where there’s a giant lack of healthcare, even way worse than in the West.
Mark Dlugoss:
Well, artificial intelligence has evolved a lot over the last 10 years. Ophthalmology has been really in the forefront of it adaptation, mostly to detect, diagnose, and treat and monitor eye disease. How is AI applied to ophthalmology specifically and what disease states are benefiting from its use?
Michael Abramoff, MD:
It’s important to realize that this is being used for real on our patients right now, every day. It’s not something that is going to happen in the future or it has potential, or whatever. No. People are being diagnosed by an autonomous AI every day. Hundreds of thousands of them. That’s important to realize. The FDA was weighed in, they know how to regulate it. CMS has reimbursed it at the national level, there is CPT coding. It can close care gaps for MIPS and HEDIS. It fits into failure based care. All the answers that you might have are of how are we going to use it or how are we going to get reimbursed, or where is it going to go, they have already been answered for a very specific AI. For, like you mentioned, diabetic retinopathy and diabetic macular edema.
Mark Dlugoss:
Moving forward, what other potential applications for artificial intelligence in ophthalmology?
Michael Abramoff, MD:
I mentioned this lack of productivity, access, lack of clinicians, quality of care. I think that’s where you see AI creators, but all the other stakeholders in healthcare focusing. That means on, for example glaucoma, that means for example AMD, where there’s a lack of access to high quality care and patients are essentially not being seen by eye care at all, so they don’t even know that they have a disease that can and should be treated. I think that’s where it starts.
There’s a whole different role for AI, which is support of clinicians, like me and my colleagues, where the AI is more assistive and helps you make maybe a faster decision or high quality decision, but you then need to prove that AI actually can do that.
Mark Dlugoss:
You actually have a balance, as a retinal specialist and an also as a member of industry. What are the challenges facing both industry and clinicians regarding the advancement of AI beyond what we’re doing with it now?
Michael Abramoff, MD:
I think it’s ultimately about patient outcome. I think we also always need to put patient outcomes first and work our way backwards from there. Now you look at oh, how can you improve the patient outcome for diabetic retinopathy? What is the process that gets the best results for the patient? What role, if any, should AI play in that? Then develop your AI from there.
What you often see is, instead, that people say, “Well, I made this cool AI. Let’s look for trying to fit a square peg into a round hole.” Then you get glamour AI, which is what I call AIs that are really cool technologically, from an engineering perspective, but don’t do much or anything rather for patient outcome. We need less of glamour AI and more of AI that actually moves the needle on patient outcome, health equity, cost, et cetera.
Mark Dlugoss:
That’s basically the key then, patient outcomes.
Michael Abramoff, MD:
I think that is key to gain and retain trust in AI, all stakeholders need to be on board. It includes patients, patient organizations, physicians and other providers, payers, ethicists, regulators, even investors. They all need to be comfortable with where it’s going. Even one of them saying no, it stops it. You can go through FDA and do everything right, and if you don’t get reimbursement, you saw it repairs therapeutics, had everything covered except reimbursement. There are randomized clinical trials, improved patient outcome, FDA approval, and they died last year, they went bankrupt because they couldn’t get reimbursement. Everyone needs to be aligned. To do that, you better have an ethical framework to build from, and better make sure that you create and retain a trust.
Part of that is making clear why it’s worth using an AI. If it improves patient outcome, well that’s more obvious than if you say, “Well, I just think it’s cool to use this AI.”
Mark Dlugoss:
You mentioned that everybody has to be on the same page. Do you feel we’re on the same page right now? In terms of ophthalmologists.
Michael Abramoff, MD:
Definitely for the autonomous AI limited score for diabetic retinopathy and diabetic macular edema, yes. It wouldn’t have happened otherwise.
You may recall, we were talking about at Ophthalmology Times before, my nickname is The Retinator, the Terminator of the retina, and it was front page news now 30 years ago. That just shows you that if you don’t get buy in from, for example, ophthalmologists and the retinal specialists, they will push back. That’s not a good thing for anyone. I think it’s key that you gain the trust of everyone.
That’s definitely the case for diabetic retinopathy. That’s why we have a CPT code, thanks to the American Academy of Ophthalmology, the AMA. That’s why there’s CMS and Medicare reimbursement, thanks to CMS. There’s regulatory approval and a path to how do measure that this AI is safe, how do you measure that this AI is effective and equitable. All these things can be measured now, thanks for FDA and other regulators. There’s this whole ethicist way, or ethical way to weigh in.
I do want to call the collaborative community for ophthalmic imaging. That is this, what you could call something that came out of FDA, where all these different stakeholders come together, we have an annual conference that’s very where well received, where we discuss these issues. What is the best reference standard to compare an AI to so you know it’s safe or not? How do you set the threshold for sensitivity or specificity? How do you evaluate explainability, racial bias, equity, et cetera? There’s so many things that we’re trying to continue to define and that is really where all stakeholders come together.
Mark Dlugoss:
Now, as the artificial intelligence continues to grow, what are the ethical concerns especially among ophthalmic clinicians? Where and with whom does the ethical responsibility lie?
Well, I think ultimately it has to come from those who make the claims about specific ethical things.
Let me tell you how we did it. We built an ethical framework for AI starting from three principles. It’s really interesting because these principles are so old. They are patient benefits, or lack of it. Their equity, or balance, or treating everyone equality. And there’s patient autonomy. These are so old, they’re thousands of years old. You see them in every culture. In some way, Confucius spoke about patient benefit, and Plato spoke about it, Mencius spoke about it. These are very old, shared across cultures.
We don’t always agree on exactly how much of each, because you need to find a balance. Let’s say you want more patient autonomy, then you have to accept that maybe the outcomes are not as good. Telling people exactly what to do when you know what they should be doing may lead to better outcomes, but that of course decreases patient autonomy. You have to find the balance between these three, and maybe more ethical principles.
But then, the next step is that you start to say well, can we actually measure how much we need of each of these ethical principles? Because you can tell an engineering, or a developer, or an AI creator, “Be ethical.” They will say, “Duh, I’m ethical.” Everyone says this. But you actually need to be able to measure it, to be able to compare and see how much you fit into these ethical principles. I think that’s key. Ultimately, that’s on the AI creator to show and claim that, “I meet the equity principles so much, there’s lack of bias there. I meet the patient benefit criteria and ethical principles so much.”
Then there’s a role for regulators to understand why the regulation should be based on ethics, and FDA gets that now. Even payers like CMS now are basing more and more decisions on these same ethical principles. For example, equity.
I think there’s a role for every one of us, but ultimately the creator says, “I think this AI that I created makes patient better off, better patient outcome, and I think it’s not biased, and it does not do other harm in terms of data usage, or costs, or maybe harming some patients.” So for all of us, but ultimately the claims come from the creator.
Mark Dlugoss:
Well, artificial intelligence obviously is already revolutionizing the field of ophthalmology. However, there are concerns how AI could affect the future of the eye care profession in general. Can you address this point and what is the reality of it all?
Michael Abramoff, MD:
I think The Retinator happening was really instructive there, because if you think about it … If you’re a clinician and you think about your own patient, and you think about there’s now this autonomous AI which does what I do, it doesn’t, but you can imagine that. Now you worry about what about me, is my job at stake?
What you see instead happening is that so many patients do not get any care, we never even think about them because we don’t see them. Now this AI has expanded our footprint as a profession, creating more health equity because now everyone can have access to high quality eye exam, in this case for diabetic retinopathy.
Yes, you can have these fears. There’s other fears. What is my role going to be? Should I focus more on procedures rather than the diagnosis? But ultimately, we’re about getting patients better outcome, more health equity, better access and AI is helping us do that.
Mark Dlugoss:
Let’s look to the future a little bit. What would be the future clinical practice, how would it look like as we move forward and incorporate more and more artificial intelligent applications into the ophthalmic practice?
Michael Abramoff, MD:
As AI becomes and scales rapidly, as you saw from the recent New England Journal of Medicine article, AI and ophthalmology, and specifically autonomous AI for the diabetic retinopathy, is the fastest growing of any AI application anywhere. By the way, getting back to something you said earlier, ophthalmology is the leader in AI in healthcare, but then healthcare is the leader in autonomous AI. You cannot buy a self-driving car. This very gnarly, complicated healthcare system of ours was the first in the world to do this.
Mark Dlugoss:
Right.
Michael Abramoff, MD:
Thanks to the ethical framework and widespread support.
But yeah, going back to the role of the clinicians, A, we will have to deal in terms of becoming more productive with all these extra patients that will now be … Most patients do not have diabetic retinopathy when you diagnose them with an AI, but some do and they need our care. So far, right now, no robot is doing injections or doing laser. They will need us for that, and then be better become more efficient so we can take care of all these patients.
Then we better start become more adept in how to use all this technology in the right way, so that we don’t harm our patients. We’d rather benefit them. Look exactly at that role, what that role should be, what the technology should do so it’s best for the patient. Not just willy-nilly use technology because it’s technologically cool.
The scribes, there’s all this new ChatGPT type scribe AI. Be very careful that it doesn’t build in equity issues that can be avoided if you do it right. There’s a lot to be done, in terms of gaining trust, validating, defining what it does, defining what we accept in terms of error rate, or whatever. Very little of that has yet been done and we will need to do it to be able to say, “Yeah, I feel comfortable using this AI on my patients.” Because don’t forget, there’s a lot of regulators and other instances looking over our shoulders in how we’re using it, and we need to show that we’re doing it the right way not the wrong way.
Mark Dlugoss:
Okay. We just got finished talking about the future of artificial intelligent, in terms of the practice. What about the future of ophthalmic surgery? Specifically, as you mentioned a retina, obviously you don’t have any robots yet, doing anything in terms of puckers, or macular holes or anything. But how do you see artificial intelligent moving into the clinic or the OR say, and helping the physician perform better outcomes?
Michael Abramoff, MD:
I think we already are seeing it, especially in cornea, with corneal surgeries. BRK, LASIK. So much of that is already automated. It becomes more and more important for the clinician to know exactly which tool is right for which patient, when to use it and when not to use it, when to intervene when it goes south. It puts higher demands on our skillset. We need to become better to be able to use these very powerful tools for the betterment of our patients, in addition to being able to deal with all these extra patients that need our care that previously were just going blind without us even being aware of it.
I think, in terms of procedures, you’re starting to see more and more of that. There’s a lot being done. Don’t forget that surgery may not be the only way to remove a cataract. There might actually be a topical or oral agent. We seem to be more worried about robots taking over cataract surgery than drops or an oral medication eliminating the need for cataract surgery in the first place. There’s so many avenues. Again, think back to what is best for the patient. What is best for the patient is to have a very safe procedure for seeing better, including removal of the cataract. Whatever gets them there the soonest, the safest, ideally the most affordable way, that’s what best for the patient, and therefore it’s best for us.
Mark Dlugoss:
I don’t want to get too philosophical here, but doing all my research for my interview here. Do you see ophthalmology currently arriving at a fork in the road when it comes to artificial intelligent? What I mean by that, you have one path, there is the physician based practice. On the other, there is the attraction and adaptation of AI. Or, is there a middle of the road option that focuses as a fusion of traditional personal care with technology and that drives the personalization of eye care? Based on your knowledge and expertise, where’s ophthalmology headed?
Michael Abramoff, MD:
Procedure medicine is such an interesting term because ultimately, the reason we can do successful surgery or even successful treatment is because, in many ways, people are very alike. If the lens is one time before the iris, and now after the iris the next time, how can you do a successful cataract surgery? No, it’s because people are so similar, not individualized, that we can even do surgery. It is key to realize that and you see the same with new drugs or new treatments, we can not only prove that these are safe and effective by averaging a large number of patients. I’m not defining large right now.
On one hand, you can never fully individualize it because if you have never used the treatment, and you do procedure medicine on the first patient, you cannot even know it’s safe at all. How do you find the middle ground between where you only think about millions of patients and what is best for the average versus what is best for the individual patient? Finding that balance is key. That is neither of the solutions you said, it’s sort of in between. You use the knowledge that you gain from people being more similar than otherwise, in terms of their response to treatment or their outcomes, and then you use that as much as you can for that individual patient because yes indeed, everyone is a little bit different and a little bit unique.
You need to understand both of these principles and need to be applied. There’s many rare diseases, orphan diseases, that we will never be able to have a large enough trial sample size to even know whether new treatments are safe. How do you apply a treatment that’s proven for a different disease to a new disease where there will never be enough scientific evidence? Well that’s because you draw parallels with the disease mechanisms, and experience, et cetera. That will remain, coupling the knowledge about large group of patients on average and the individual patient, and marrying the two. That’s a trick that I don’t see changing very much.
Mark Dlugoss:
Well, we’ve covered a lot of information here in our short time on artificial intelligent. Are there any points of discussion I may have missed in asking you questions or is there anything you’d like to add regarding artificial intelligence as its evolution into ophthalmology?
Michael Abramoff, MD:
I’m so proud, again pointing back to this New England Journal of Medicine AI article a few weeks ago, that we’re leading the way in ophthalmology and retina, what has been achieved already because it’s all there. Regulatory considerations, reimbursement, failure based care, all the aspects, all the ethical aspects have all been addressed, continue to need to be addressed. There’s a lot of work ahead. But it can be done, we’ve shown it, and we can do it the right way. Rather than harming patients, this AI is benefiting tremendously in terms of health equity, patient outcomes, clinician and retina specialist productivity, job satisfaction for retina specialists. I presented that recently at Macular Society.
It’s going so well. We need to keep doing it right to not stop this advance and continue to benefit our patients.
Mark Dlugoss:
Thank you, Dr. Abramoff.
Well, that concludes today’s Ophthalmic Project. I want to thank Dr. Abramoff in spending some time with me outlining the current status and the future of artificial intelligence and ophthalmology. I also want to thank you, the viewers, for watching. I hope you’ll join us for the next edition of The Ophthalmic Project. Have a great day.