This is a supplement to EyeWorld Magazine.
Issue link: https://supplements.eyeworld.org/i/1538645
T he 2025 ASCRS Annual Meeting featured two symposia presented by the ASCRS Digital Clinical Committee and moderated by John Hovanesian, MD, chair of the committee. 'Practical AI Tools that Can Enhance Your Practice Today' The first symposium featured presentations from committee members and companies in this space. It was sponsored by Alcon, Amaros, AVTR Med, and Zeiss. Demystifying AI nomenclature Grayson W. Armstrong, MD, MPH, said there are a lot of different AI companies coming into ophthalmology and sim- plifying practices and surgeries. "But if we don't know the terminology, it's going to be harder for us to vet that tech- nology for our practices, it will be harder when interfacing with those companies to know what's best for our practices and patients, and it will be harder to educate our patients about them," he said. Some of the AI tools that are entering clinics include autonomous diabetic retinopathy screening, virtual AI-based scribes, and automated inbox triage. Knowing the terminology empowers you to assess the clinical value of new technology; communicate confident- ly with vendors, IT teams, and patients; and advocate for responsible tech adoption, Dr. Armstrong said. He gave an overview of definitions of AI, noting that AI in general is any computer system that mimics human intel- ligence. A subset of that is machine learning, which learns patterns from data, and within that is deep learning, a type of machine learning using layered neural networks. There are multiple common AI architectures in ophthal- mology, Dr. Armstrong said. Convolutional neural networks take individual pixels or groups of pixels in image analysis and can tell if it shows disease, such as diabetic retinopathy. Meanwhile, large language models are good for text-based activities. Random forest models are decision-tree prediction models that can tell you the risk of something. Other performance metrics that are important to know include accuracy, sensitivity/specificity, confusion matrices, positive predictive value and negative predictive value, AUC/AUROC, and F1 score. All these metrics are different ways of showing how the algorithm will function in your patient population and how accurate the algorithm is at detecting normal vs. abnormal cases. Dr. Armstrong also noted several questions to ask before deploying an AI tool in your clinic: 1. Was it validated in a population like mine? 2. Does the research show a high level of AI accuracy? 3. Does the algorithm improve care, efficiency, or outcomes? AI scribing and telephone management Robert Chang, MD, presented on AI scribing and telephone management, including applied uses of AI. He started by showing a speaking AI avatar, which he said can be created easily and inexpensively. AI agents are the next evolution of large language models, Dr. Chang said. These agents can execute tasks in sequence and with memory and are not limited to just what has been trained in the model. Virtual scribes can create a summary note from listening to a doctor-patient conversation, suggest ICD-10 and CPT codes based on a visit, pre-chart test orders and pre- scription refills for physicians to sign, create an after-visit summary geared for patients, and pre-write answers to patient messages. Dr. Chang said he has a virtual scribe built into his EHR software. AI can insert key points from the patient conver- sation into a custom template, thus displaying visual acuity and the slit lamp exam in the style you like. AI can act like a receptionist using a phone agent AI, he said. You can customize the personality of the replies, including changing primary language, building different AI terminology AI algorithm: A decision-making computer-based "recipe" Training data: Data that the model learns from Ground truth: The "correct" answer Overfitting: Algorithm is "too specific" to training data Generalizability: Not overfit, and algorithm works beyond test training set Explainability: Can humans understand the decision? Black box: Opposite of explainable; opaque ASCRS DIGITAL CLINICAL COMMITTEE SYMPOSIA AT THE 2025 ASCRS ANNUAL MEETING