Improving Fairness and Reducing Bias in Artificial Intelligence Algorithms for Glaucoma - Project Summary/Abstract: Glaucoma is the leading cause of irreversible blindness, affecting over 60 million people worldwide. Glaucoma patients vary widely in their presentation, with some retaining long-term disease stability, and others progressing quickly to vision loss. Recent advances in artificial intelligence (AI) for glaucoma have enabled integration of the rich and complex EHR data into algorithms that predict which patients will have progressive glaucoma, potentially enabling personalized treatments plans that prevent vision loss and reduce the overall burden of glaucoma care. In the rapidly evolving landscape of healthcare artificial intelligence (AI), there is a pressing need to prevent issues that may inadvertently worsen healthcare outcomes, if AI algorithms do not work well across different phenotypes of patients. For example, when the population used to train AI algorithms is dissimilar to populations where AI algorithms will be deployed, the generalizability of the AI algorithms across those populations is key to their effectiveness. There is an unexplored frontier in investigating generalizability in ophthalmology AI algorithms, largely due to the scarcity of large-scale multicenter ophthalmology datasets. The newly established multicenter Sight Outcomes Research Collaborative (SOURCE) registry, comprising data from 23 centers across the U.S., offers an unprecedented opportunity to assess AI algorithms in real-world scenarios and across different populations. The overall goal of this project is to develop generalizable prediction algorithms for glaucoma progression using multicenter scale EHR data. Aim 1 will compare established algorithm training methods aimed at standardizing performance between groups to investigate the tradeoff between overall algorithm performance and performance across subgroups, measured by equalized odds. Aim 2 will leverage the multicenter nature of SOURCE to train and test glaucoma prediction models across sites with highly varied population characteristics, investigating the impact of site and population differences on algorithm performance. Aim 3 will evaluate the use of novel group distributionally robust optimization methods to enhance generalizability in glaucoma prediction algorithms. Throughout this innovative project, we will use state-of-the-art AI methods for training generalizable algorithms to develop our glaucoma prediction algorithms in a large new multi-center ophthalmology EHR registry. Completion of these aims will mark the first systematic studies of generalizability for electronic health record prediction models in ophthalmology. The insights gained will not only optimize generalizability and performance in glaucoma AI but also extend to broader applications across many medical disciplines. By unveiling the intricacies of how AI algorithms perform when trained and tested in a variety of populations, this research is poised to significantly improve health outcomes for ophthalmology patients and serve as a guiding beacon for responsible AI implementation in healthcare.
