Eyeworld

September 2018 • Ophthalmology Business 15 pathology, AI will almost certainly create a global marketplace that will facilitate access to quality care from which perhaps half the world's population is deprived. The extent to which that global marketplace might disrupt local practices in industrially advanced countries remains to be seen. But with AI technologies ad- vancing at a rapid rate such that their proceeding to FDA clearance in the U.S. will not be far off, it behooves everyone with a stake in the outcome to learn and estimate the quality and scope of AI's clinical applications so as to plan their business futures wisely. OB References 1. Allain J. From Jeopardy! To Jaundice: The Medical Liability Implications of Dr. Watson and Other Artificial Intelligence Systems. Louisiana Law Review. 2013;73:973–1049. 2. Jha S, Topol EJ. Adapting to artificial intelligence: radiologists and pathologists as in- formation specialists. JAMA. 2016;316:2353– 2354. 3. Gulshan V, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photo- graphs. JAMA. 2016;316:2402–2410. 4. Beam AL, Kohane IS. Translating artifi- cial intelligence into clinical care. JAMA. 206;316:2368–2369. 5. Wong TY, Bressler NM. Artificial intelli- gence with deep learning technology looks into diabetic retinopathy screening. JAMA. 2016;316:2366–2367. 6. Wu X, et al. Prevalence and epidemiolog- ical characteristics of congenital cataract: a systematic review and meta-analysis. Sci Rep. 2016;6:28564. 7. Wang A, Shen D. Computational medicine: a cybernetic eye for rare disease. Nature Biomedical Engineering. 2017;1:0032. 8. Kantarjian H, Yu PP. Artificial intelligence, big data, and cancer. JAMA Oncol. 2015;1:573–4. and other image-dependent practic- es might just send their images off to centralized reading sites that will interpret the images for a very low cost. 5 Doing so might significantly reduce costs for an ophthalmology practice or, alternatively, free up clinicians and staff to perform more revenue-enhancing tasks. Such "read- ing centers" may well become com- monplace for image interpretation in the years to come, as their access to massive registries, high output deliv- ery that operates 24 hours a day, and associated high levels of accuracy and reliability will enhance their market- place appeal. Regardless of whether we are talking about screening and treat- ment recommendations for common or rare eye conditions, however, any ophthalmology practice is obviously going to scrutinize the quality of AI technologies before buying and using them. An important starting point for these algorithms is the extent to which their data training sets are adequate given the disease variations that patients will present. A training set may require hundreds of thou- sands if not millions of images or patient records to ensure an accept- able level of reliability and accuracy. In turn, that need will likely result in new, entrepreneurially based business models that will negotiate with clin- ics and hospitals to acquire and test AI training materials so as to ensure product quality. Additionally, note that Gulshan's study relied on the interpretations of practicing ophthal- mologists to build and inform his training set of images, which means that the quality of the AI's ultimate outputs is only as good as the quality of their inputs. As Gulshan and his colleagues admitted, "This means the algorithm may not perform as well as images with subtle findings that a majority of ophthalmologists would not identify." Perhaps more than anything, though, these technologies will open up business opportunities for image- reliant practices throughout the world. When we think of the busi- ness of ophthalmology, radiology, or These kinds of technologies and their benefits will only continue to improve. Some of these benefits will include developing large disease or illness registries whereby epide- miologic trends can be followed; detecting new disease etiologies or correlations; identifying beneficial therapies custom-tailored for partic- ular patients; tracking differential outcomes; developing new standards of care; comparing and evaluating existing treatment modalities; and tracking treatment trends and their outcomes. 8 But given the two studies mentioned, how might the clinical and business practices of ophthal- mology change? Considering the Gulshan study on screening for diabetic retinopathy, we should immediately note that the training data set was specialized for that disease and not for glaucoma or age-related macular degeneration— which diabetic retinopathy screening programs would normally include. Consequently, it remains to be seen how such an AI application would be integrated with the other diag- nostic tests that a clinician would normally perform. And as Gulshan et al. admit, "this algorithm is not a replacement for a comprehensive eye examination, which has many components such as visual acuity, refraction, slitlamp examinations, and eye pressure measurements." An- other problem is the degree to which such a technology, once incorporated into a retinal camera, would simply be relied on by the ophthalmologist whose trust might become so great that he or she no longer reviews the AI's image interpretation. 5 Many think that such a degree of explicit reliance on the technology with little if any human oversight will almost certainly happen over time, as phy- sicians and office staff scramble to meet their productivity targets. If so and the AI system errs, however, how will liability be apportioned in the event of an adverse occurrence and a lawsuit? 1 A third issue that speaks to both of the studies mentioned involves how the business of ophthalmology Dr. Banja is a med- ical ethicist at the Center for Ethics at Emory University, Atlanta. He can be contacted jbanja@ emory.edu. About the author

• Blank Page
• Blank Page (1)