Jorge Cuadros on AI Detection of Diabetic Retinopathy and Beyond

David Klonoff: 0:14

Hello, welcome to Diabetes Technology Report podcast. I'm David Klonoff, endocrinologist at Sutter Health and UCSF, and we have a special guest today. My co-moderator, David Kerr, will introduce himself and will introduce our guest.

David Kerr: 0:32

Welcome everyone and it's great that you can join us, and it's a real pleasure to welcome Jorge Cuadros. Jorge, we know you very well, but I just want to make sure that all our listeners are aware of your background and where are you working at the moment.

Jorge Cuadros: 0:49

Sure. So my name is Jorge Cuadros and I'm the director of Informatics Research in the Department of Optometry and Vision Science at University of California, Berkeley, and I'm also CEO of IPACS, incorporated, a diabetic retinopathy screening program that's been around since 2003.

David Kerr: 1:13

Great, well, welcome. When I was doing my training, I used to do a rational clinic with an ophthalmologist and he was so passionate he said that the eye was the gateway to the soul. Why did you, or how did you, come to get interested in retinopathy and all things related to the eye?

Jorge Cuadros: 1:35

Yeah, I was interested in telemedicine since 1993, but in particular with diabetic retinopathy, because even though there's effective treatments, it remains the main cause of blindness among working-age adults, and it's because of a lack of appropriate detection and follow-up into the ophthalmologist's office for treatment. So that's what really got me interested in it. And then over the years, through the work that we've done with artificial intelligence, we're finding that in fact, yes, it truly is the window to so many other things that are happening in the body. So even things that humans can't see in a retinal image can be detected with artificial intelligence and some of the new discoveries and innovations that we're coming across.

David Klonoff: 2:39

Jorge, could you explain what is artificial intelligence and how is it applied to your work? Screening for diabetic retinopathy.

Jorge Cuadros: 2:47

So artificial intelligence has been around for a long, long time, but because of developments in computing power and resources that have become available, it's now becoming much more possible to use different artificial intelligence methods, particular neural networks, artificial neural networks in order to detect and diagnose diseases, one of them being diabetic retinopathy Retinopathy, in fact, that was the first disease that was used, that where artificial intelligence was able to detect diseases.

David Klonoff: 3:33

Jorge, how accurate is artificial intelligence when it comes to diagnosing retinopathy?

Jorge Cuadros: 3:40

So back in 2015, with the help of the California Healthcare Foundation, we did a competition. We put 50,000 retinal images with diabetic retinopathy, some without to see how accurately people around the world, teams around the world could develop their artificial intelligence algorithms. Within three months, hundreds of teams were already outperforming humans in being able to detect diabetic retinopathy. So the sensitivity and specificity for AI-driven programs is better than humans looking at retinal images. The only catch is that they sometimes require better quality images than what's generally available in a real-world setting, and that has been something that's restricted just autonomous artificial intelligence in doing large-scale screening programs.

David Kerr: 4:49

So, going forward, your focus has been on diabetes and retinopathy and the different kinds of retinopathy and loss of vision, but are you starting to see insights into other disease processes, other diseases where the retina can be a true gateway to the soul?

Jorge Cuadros: 5:10

Yes, in fact there's been some new algorithms being developed that were trained with electronic medical record data to be able to partition retinas based on other diseases, such as kidney disease or liver disease, with really great results. So when a human is looking at a retina, it's difficult to know if there is active kidney disease or to even be able to stage the kidney disease, but the algorithms are able to do that effectively, very accurately. So cardiovascular disease, liver disease, cognitive disorders there's a rich area of research where algorithms are being developed to be able to detect those conditions just from a retinal image alone.

David Kerr: 6:11

That is astonishing. So are we getting into Alzheimer's disease detection and things like that?

Jorge Cuadros: 6:20

That's right, and not only that. So looking for these biomarkers in the retina is pretty well established. But what about other parts of the eye? So we've been working. In fact, Google published an article recently about external images, images of the outside of the eye, being able to detect liver disease and kidney disease and cardiovascular disease. All of those. And sure enough, there are unknown biomarkers, unknown to us, to humans, that are used by the algorithms to be able to detect these diseases. So you could imagine that maybe in a few years, just with a smartphone taking a picture of the front of the eye, you'll be able to detect the need for seeing a physician.

David Klonoff: 7:21

Jorge, you run an organization, Ipacs. Could you explain what is Ipacs?

Jorge Cuadros: 7:26

So Ipacs started as a telemedicine-based service where retinal cameras are put into primary care organizations. Images are captured of patients Up until now it's been mostly diabetic patients, but that's been broadening For the purpose of early detection and triage of patients that have site-threatening conditions. Ipacs is a turnkey service for that, providing the cameras, the expertise, the ophthalmologist, optometrists that read the images and all of the training and quality assurance that the primary care clinics would need. We've also been involved in a lot of research to develop these algorithms and looking for ways where technology can improve access to care, can reduce healthcare disparities, especially in vulnerable populations.

David Klonoff: 8:29

I imagine part of your approach is to screen people without having to have a physician at the site. How easy or likely is it that a person can be screened without having to have their pupils dilated?

Jorge Cuadros: 8:43

Right. So almost all of the images are captured without pupil dilation. So the technology is continuing to advance. It's getting easier and easier to capture images. We always have to keep an eye out for image quality and to assure that we're not just producing images that aren't able to detect the things that either humans or the algorithms can detect. So quality assurance is also is always a big piece of this. But the technology is improving, getting smaller, getting more portable and more accessible, less costly. So I would imagine in a few years it'll be readily available anywhere, maybe at home. One interesting area of development is an autonomous image capture. So essentially you just put a kiosk, a patient walks up, gets their picture captured. You know the retina captured. It's getting to be that easy.

David Kerr: 9:50

One of the features, just going back to diabetes in the eye, is that we know that retinopathy is not rigid, that it's a fluid disease process in the eye. Some progress, some stay the same, some regress. In terms of what you're doing, are we now able to stratify people into risk of progression and then change the frequency of the screening and then suggest, you know, different interventions that might be helpful in terms of risk reduction? Are we at that sort of place already? I would say so.

Jorge Cuadros: 10:30

So we can track patients over time, and it's always interesting to see how some patients who have not controlled their blood sugar very well don't progress. They have very little retinopathy even after years, and then others are much more susceptible to changes in their blood sugar and up until now, just based on an image alone, humans have not been able to detect that, but algorithms have. So that's another area of research is to be able to predict the rate of change in retinopathy just based on a single retinal image.

David Klonoff: 11:09

Jorge, it's important to diagnose retinopathy so people are eligible for screening. How successful are your programs and others' programs in actually convincing people to go in and receive treatment such as laser treatment?

Jorge Cuadros: 11:24

So glad you brought up that point and it's the most important part of the research that we're doing is how to close the loop and ensure that someone who's tested and found to have site-threatening condition actually gets treated, and for that it's a much more complex communication system with the specialists and the networks and the eligibility verification and finding out well how long can they wait being able to stage the referrals. There's a lot of social determinants of health to overcome and just logistics in general, and it's encouraging that the immediate results of an algorithm really help a lot in getting patients into care. If you're able to do the test and immediately get the results, it's much more likely that you can engage with the patients in order for them to take the next step in their treatment.

David Klonoff: 12:34

And with AI you can get the results immediately, Whereas with typical telemedicine systems.

Jorge Cuadros: 12:40

you have to wait for the consultant to become available, and sometimes by that time the patient's already gone and thinking about something else.

David Klonoff: 12:49

Well, Jorge, thank you for speaking to us from UC Berkeley. This has been an enlightening podcast. This completes the podcast For listeners. You can find Diabetes Technology Report on Spotify, the Apple Store, Diabetes Technology website, and for me and Dr David Kerr and Dr Jorge Cuadros. Thank you for listening and we'll catch you at the next Diabetes Technology Report. Bye-bye, Bye-bye.