EyeWorld Asia-Pacific December 2020 Issue

DEVICES 52 EWAP DECEMBER 2020 get better and better with time and larger amounts of data,” she said, but she added that machine learning is “like teaching a child.” You need to keep reinforcing these lessons so the machine learns better, and that will come with time and repetition, Dr. Habash said. Dr. Trubnik also noted potential limitations, particularly being careful that you don’t have confirmation bias with such a large sample size of data. She stressed the difference between association and causation. Just because you find the association doesn’t mean it leads to something and you have to act upon it, Dr. Trubnik said. When talking about deep learning, Dr. Trubnik said there is a “black box” for how AI comes up with a diagnosis; there may be questions from patients on how the diagnosis was determined. “You develop a relationship with the patient and explain things to them, but if the diagnosis is coming from a machine, they may question how you arrived at the diagnosis,” she said. Dr. Trubnik added that for big data to be useful, a lot of images are needed, and sometimes there aren’t enough to yield accurate information. Also, when sharing information, she said, sometimes it’s hard to know if the data is all equal. Additionally, she said that rare diseases may be difficult to diagnose with AI because there is not enough data to recognize them. Specifically relating to glaucoma, Dr. Trubnik said that arriving at the same definition of glaucoma may be a challenge. “The other thing is if the AI is screening fundoscopy or OCTs, usually it’s utilizing a ‘binary classification,’ so it will say ‘yes glaucoma’ or ‘no glaucoma,’ but when I examine my patient, I can also pick up cataract, epiretinal membrane, etc.,” Dr. Trubnik said. “So far, they haven’t been able to do that accurately with AI because when they incorporate multiple variables, it’s not as accurate at predicting.” Going back to the gym analogy, Dr. Habash said, “Technology helps us lift the weight, but we’re still the ones doing the work. AI and telemedicine won’t replace us, but they will augment us.” EWAP Editors’ note: Dr. Habash practices at Bascom Palmer Eye Institute, Miami, Florida, and has interests with Microsoft. Dr. Trubnik practices at Ophthalmic Consultants of Long Island, Lynbrook, New York, and declared no conflicting interests. Artificial intelligence to improve cataract surgery outcomes - from page 50 tomography, and postop outcomes. Dr. Solomon said the collection of this data is beginning with programs like Veracity Surgical (Carl Zeiss Meditec). While Veracity has tens of thousands of data points to enhance surgical decision making through its program, Dr. Solomon said when it gets into the millions, there will be greater ability to make better decisions. “How do we get there? It has to be more widely distributed, more widely accepted, offices have to use it, etc. For that to occur, more offices have to accept it, perhaps the process has to become more streamlined and more intuitive and user friendly,” he said. “Do I think we’ll get there with Veracity? Absolutely. Will there be other products beside Veracity? Absolutely.” What’s needed for AI to benefit cataract surgery Further refinements are needed to best apply AI for IOL power calculations, Dr. Hill said, explaining that the main limitation is the accuracy of metrics used for model fitting. He said preop keratometry and the postop refraction are two areas with the greatest variability. He added that as technology used to collect these and other measurements becomes more accurate, artificial intelligence models will offer more predictable outcomes. Dr. Solomon said for true potential to be realized datasets also need to grow. “We have to get rid of the data that’s in silos. Every office has data, whether in a paper chart, an EMR chart, sitting in your biometer, topographer, or OCT machine, we all have data on our own patients. Most of us don’t have access to our own data and most of us don’t have access to other people’s data, and the only way we’re truly going to have big data is if we can break down these silos and allow us to put this data in a cloud so we can share this information so we can all benefit from it,” Dr. Solomon said. Dr. Solomon also cautioned against the notion that AI is a “computer telling me what to do.” “Nothing could be further from the truth,” Dr. Solomon said. “All the computer is doing is helping to assimilate a bunch of information and taking the doctor’s logic and the doctor’s own preferences and helping them assimilate it to simplify the decision making they would ultimately arrive at, if they had the right information. But how would an individual practitioner be able to assimilate information from a million cases? You wouldn’t.” That’s where AI comes in. EWAP Reference 1. Goh J, et al. Artificial intelligence for cataract detection and management. Asia Pac J Ophthalmol . 2020;9:88–95. Editors’ note: Dr. Hill practices at East Valley Ophthalmology, Mesa, Arizona, and has interests with Haag-Streit Diagnostics. Dr. Solomon practices at Carolina Eyecare, Mount Pleasant, South Carolina, and has interests with Carl Zeiss Meditec.