Improving Fairness and Reducing Bias in Artificial Intelligence Algorithms for Glaucoma - Glaucoma is the leading cause of irreversible blindness worldwide. Some glaucoma patients can remain stable with good vision for years, while others rapidly progress to vision loss, and it can be difficult for clinicians to predict a patient’s clinical trajectory. With the aid of personalized prediction algorithms for glaucoma, clinicians could customize patients’ therapeutic goals to intervene in high-risk patients before vision is irreversibly lost, or relax the burden of testing in stable patients. One of the most promising methods for identifying patients at high risk for glaucoma progression is by leveraging artificial intelligence techniques on electronic health records (EHR) to build high-performance predictive algorithms, an area which Dr Wang’s research group has pioneered. However, a critical barrier in implementation of prediction algorithms lies in understanding the impact of these algorithms on health equity, as algorithmic bias has been shown to particularly affect minoritized groups and potentially deepen health disparities across many medical domains. In the U.S., glaucoma disproportionately impacts Black and Hispanic patients, further intensifying the urgency of addressing biases in artificial intelligence (AI) prediction models for glaucoma. With the recent establishment of the multicenter Sight Outcomes Research Collaborative (SOURCE) registry, we now have the opportunity to evaluate the fairness and generalizability of our AI algorithms for glaucoma in EHR data from 17 diverse centers across the United States. The overall goal of this project is to develop fair and generalizable prediction algorithms for glaucoma progression using multicenter EHR data. We hypothesize that there exist racial/ethnic disparities in the performance of our glaucoma AI algorithms which depend on the characteristics of the populations that the algorithms are trained and deployed on, and that we can improve the fairness of our AI algorithms by using fairness-aware training techniques. To investigate these hypotheses, we propose three aims. Aim 1 will compare different “fairness-aware” algorithm training methods for bias mitigation to improve fairness for glaucoma prediction models. Aim 2 will leverage the multicenter nature of SOURCE to train and test glaucoma prediction methods across sites with highly varied population characteristics, investigating the impact of site population differences on algorithm performance and fairness. Aim 3 will evaluate the use of novel group distributionally robust optimization methods to enhance generalizability and fairness in glaucoma prediction algorithms. Taken together, these studies will be the first systematic studies of fairness for EHR prediction models in ophthalmology, which will ultimately promote health equity for ophthalmology patients. By comparing fairness-aware and unaware training methods, evaluating generalizability across diverse independent populations, and optimizing algorithms for robustness, this research will not only address racial/ethnic disparities in glaucoma AI algorithms but also advance our understanding of fairness and generalizability in EHR models in the broader field medicine and AI.
