Eyeworld Supplements

continued on next page pathways based on scheduling vs. directions vs. reminders, etc. The workflow can be customized using existing practice information from multiple sources such as your website, allowing the AI to follow specific scripts and listen for men- tion of certain conditions. validation and quality control and can be customized for a particular surgeon or generalized for multiple surgeons by pooling data. Early results show up to 94% accuracy, Dr. Smith said. Grading lens opacity is crucial for surgical planning, and AI models can quantify nuclear sclerosis, cortical opacity, and posterior subcapsular opacity. They can suggest timing for surgery and reduce interobserver variability. All of this is hopefully going to enhance our accuracy, allowing us to have higher postop satisfaction for patients and reduce refractive surprises, Dr. Smith said. It will help with standardized grading for better surgical scheduling and support surgeons in high-volume or resource-limited settings. There are some limitations, he said, like data bias and generalizability, black box concerns in AI decision making, regulatory and medical-legal challenges, importance of human oversight, and inconclusive data. Dr. Kitchens presented on the expected increase of diabetes worldwide and how AI can help improve healthcare for these patients. Source: ASCRS Dr. Smith spoke about AI-assisted IOL calculations and grading in cataract surgery, saying that this will hopefully enhance accuracy and allow for higher postop patient satisfaction. Source: ASCRS Lens calculations and grading Ryan G. Smith, MD, spoke about AI-assisted IOL calculations and grading in cataract surgery. He mentioned the many formulas available and the measurements they require. Dr. Smith said these formulas have limitations in that they make assumptions of ocular anatomy and have reduced accuracy in extreme eyes. Newer formulas incorporate some or all standard geometric optics, ray tracing, and AI or machine learning. Dr. Smith said the first attempt of using AI in IOL power selection was in 1997, using a neural network that had 62.5% accuracy. It was cumbersome and bulky. Advanced machine learning algorithms can be utilized to create predictive models, which can be updated and improved as postoperative refractive data becomes avail- able. Current AI formula models include the Hill-RBF, Ladas Super Formula, EVO, Kane, Zeiss AI IOL Calculator, Pearl- DGS, and Hoffer QST. Some use big data and complex relationships beyond traditional formulas. The Hill-RBF uses radial basis function neural network; the Kane combines theoretical optics with AI; the Zeiss AI IOL Calculator is built into Veracity surgical planning software (Zeiss) and is trained by paraxial ray tracing generating 500,000 model eyes and crowdsourced outcomes data from Veracity users for specific IOL models. One thing on the horizon that Dr. Smith is excited about is the Precision Ladas Universal Super (PLUS) algorithm. With a training set of only a few hundred eyes, it utiliz- es ensemble learning with gradient boosting to focus on Population screening for diabetic retinopathy John Kitchens, MD, presented on the expected increase of diabetes in the worldwide population and how AI is already helping improve healthcare for these patients. He said 50% of patients who have diabetes do not undergo routine eye examinations. Barriers to care include limited access to eyecare providers, cost, and education. What makes diabetic retinopathy and screening ideal for AI is we have a large database of graded photos, Dr. Kitchens said. When you're training a deep learning or large lan- guage network to analyze, you need to have a comparison. There are three FDA-approved AI algorithms for diag- nosing diabetic retinopathy patients: IDx-DR (now known as LumineticsCore, Digital Diagnostics), EyeArt (Eyenuk),

• Cover