Achieving Model Fairness on Automatic Primary Open-angle Glaucoma Screening

点击翻译按钮获取中文摘要

Project summary/abstract In the United States, primary open-angle glaucoma (POAG) is the leading cause of blindness, especially among African and Hispanic Americans. Because visual function loss from POAG is irreversible, it is critical to estimate the risk of POAG and prevent further vision loss. Recently, there has been growing concern that the predictive model may reflect and amplify human bias and reduce the quality of their performance if used in the clinical pipeline for patient triage. Motivated by known differences in disease manifestation in patients such as sex and race/ethnicity, this study hypothesizes that algorithms trained on existing datasets will exhibit systematic biases in subpopulations. Popular approaches to remove such biases suggested that having a greater number of positive cases across demographics helped models perform better in validation. However, collecting new data often suffers from a lack of demographic representation. In response to NOT-EY-22-004 (Research Addressing Eye and Vision Health Equity/Health Disparities) and PAR-22-141 (Secondary Analysis of Existing Datasets), this project will develop and validate a new artificial intelligence approach to improve the fairness of the predictive model on POAG risk estimation without the need for demographically balanced datasets. Based on our preliminary data and our experience with an interdisciplinary team of data scientists and ophthalmologists, we plan to execute specific aims: 1) studying “algorithmic bias” in the POAG risk estimation and 2) examining the impact of “transfer bias” from the biased to the demographically balanced data. The studies proposed in this project are novel and innovative because the secondary analyses of existing data provide additional insight into POAG health disparities. Aim 1 will be the first to perform a systematic study of algorithm bias in the DL-based POAG predictive models and identify the factors contributing to model fairness. Aim 2 will be the first study to examine that bias transfer may arise in the POAG prediction setting and can occur even when the POAG dataset is explicitly de-biased. We argue that our models provide simple, interpretable, and easily checkable frameworks to allow better POAG risk estimation for protected groups. The expected outcome of this project is a holistic framework to mitigate the impacts of inequity by improving the inference performance for minorities. The success of this project will provide additional insight into health disparities of POAG risk estimation by (1) reducing clinical decisions tainted by unconscious or conscious bias, and (2) developing brand-new models that reflect learned POAG features but not patient demographic to ensure robustness across diverse populations. This project is highly feasible and potentially transformative for both data science and clinical medicine.