Eyeworld

22 Ophthalmology Business • December 2017 away. We hope that they're going to help you do a better job at what you do in diagnosing and managing patients." There are many aspects of diagnosis and management that are uniquely human, he said, like how a doctor connects with patients and gets information from them. These systems are not as good at asking patients what's wrong, what their symptoms are, what priorities they have, and they're not as good at figuring out what to do after a diag- nosis, Dr. Chiang said. "What goes into a management plan is not just algorithms, it's got to do with patients' risk aversion, their priorities, what they value. I think management is a more subjective thing. Diagnosis is more objective," Dr. Chiang explained. His hope is that doctors will embrace these technologies because "whether we like it or not, these sys- tems are coming. I think it's natural to assume that someday, more of them are going to be able to make a more accurate and reproducible diag- nosis than many doctors." OB Reference 1. Long E, et al. An artificial intelligence plat- form for the multihospital collaborative man- agement of congenital cataracts. Nat Biomed Eng. Published online January 30, 2017. 2. Abramoff M, et al. Improved automat- ed detection of diabetic retinopathy on a publicly available dataset through integration of deep learning. Invest Ophthalmol Vis Sci. 2016;57:5200–5206. Editors' note: Dr. Abramoff has finan- cial interests with IDx LLC, which commercializes diagnostic devices for diabetic retinopathy and other diseases, and Alimera Sciences (Alpharetta, Geor- gia). Dr. Chiang has financial interests with Novartis (Basel, Switzerland). Contact information Abramoff: michael-abramoff@uiowa.edu Chiang: chiangm@ohsu.edu neovascularization in the retina. With central retinal vein occlusion, there are also hemorrhages in the retina, but they're a different pattern. It's very easy to recognize that when we [as experts] are looking at a retina, we're looking at a central retinal vein occlusion, not diabetic retinopathy. But computer systems often get con- fused by that, so the scope of their capabilities is limited in this regard." In addition, he said that to make an accurate diagnosis, you have to have high quality images, and some of these systems are not as good at recognizing when they have suffi- cient quality of an image to make the right diagnosis. "Machines are not perfect. Their strengths are different. Ultimately it's doctors who take care of patients, not machines—but the machines can help," Dr. Chiang said. Assisting, not replacing Dr. Abramoff said he thinks diagnos- tic AI systems need to go where the patients are, which is primary care. "People with diabetes are man- aged there, diagnosed there, and when they're abnormal and need specialist care, they end up with a specialist," Dr. Abramoff said. "I think a lot of the routine diagnostics that we do now are going to shift to automation, to primary care, then we as clinicians will be able to do more complex treatments, more complex management. We gain time for taking care of our complex patients because we lose the routine tasks; they go to front line and there will be more complex management of gene therapy and stem cell therapy, for example. We'll need the time because there are so many patients needing it, and we can't handle them now." A lot of doctors get nervous when they see these computer sys- tems because they're worried, "Is it going to take my job away?" Dr. Chiang said, "We hope that they're not going to take your job our current practice. I bet if you look for AI in medicine, you will find a lot of it in diabetic retinopathy. It's because the classification systems and management guidelines are so well worked out." He continued: "There's a lot of scientific evidence on why we do the things we do, what to do with these types of patients. We have drugs, lasers, treatments, and we know exactly when and where to use them. There's a preferred practice pattern, recommendations on how to treat patients that are very specific. You don't see that as much elsewhere." Safe use Of course, patient safety is the most important factor in implementation. Dr. Abramoff, who is currently leading the development of guide- lines for using artificial intelligence for autonomous diagnosis in the retina, said, "If you have a diagnos- tic device, I think it's important you know what that device does. It's not enough to show this device is doing well when tested, you have to be able to say, 'Here is where it's analyzing hemorrhages and how many, here is where it detects the optic disc.' You need to be able to explain how the system does these things." Otherwise we'll lose the trust of patients, he said. "It's crucial that we keep paying attention to safety, that we explain what we're doing, why we're doing it, that we don't try to push it too fast," Dr. Abramoff said. It's important for doctors to understand the limitations of these computer systems, if we end up using them, Dr. Chiang said. For example, the diagnostic sys- tems for identifying and classifying diabetic retinopathy are often not good at recognizing when a patient has a disease other than diabetic retinopathy. Dr. Chiang continued: "With diabetic retinopathy, one of the characteristics is hemorrhages and continued from page 21

• Blank Page
• Blank Page (1)