SCH: Quantifying and mitigating demographic biases of machine learning in real world radiology

点击翻译按钮获取中文摘要

PROJECT SUMMARY (See instructions): The application of modern machine learning algorithms in radiology continues to grow, as these tools represent potential huge improvements in efficiency, accessibility and accuracy of diagnostic and screening tools. At the same time, these increasingly complex machine learning models can have biased predictions against individuals of under-represented demographic groups, potentially perpetuating pre-existing health disparities. Such fairness concerns are particularly important in public health applications that focus on large scale population-based screening, as in cancer screening for breast and lung cancer. In these settings, it is paramount to understand how often machine learning screening algorithms can be unfair and biased, and how to mitigate these disparities. This proposal will develop tools to quantify, correct, and analyze the biases of predictive algorithms in relation to different demographic groups in real world settings. In particular, we will develop analysis and algorithms to quantify the violation of fairness by a machine learning model in situations where information about the sensitive attribute itself (such as biological sex, race or age) are not directly observable, and we will provide algorithms that correct for their worst-case fairness violations. We will analyze our tools under distribution shifts, whereby differences in populations exist, as is common in large scale cancer screening programs. This project will also perform inference on the training samples and features most highly associated with fairness violations, thereby providing guidance on the development of solutions to prevent biased algorithms in the future. Our tools will be validated on a variety of large real-world radiology datasets spanning multiple imaging modalities, including general chest X-ray datasets that include lung cancer diagnoses (CheXpert and MIMIC-CXR), as well as the Emory Breast Cancer Imaging Dataset (EMBED) and the National Lung Cancer Screening Trial, evaluating and correcting disparities for predictive algorithms with respect to biological sex (where appropriate), race, and age. The results of this project will establish critical knowledge about the propensity of machine learning models for medical imaging diagnosis and cancer screening to be unfair and biased, as well as foundational tools to quantify and mitigate these biases in these potentially game-changing technologies.