Developing a Hive Learning and Datathon Supported Course on Imaging and Multimodal Data for Resource-Limited Institutions (CIMDAR-HIVE) - ABSTRACT AI tools such as deep learning models promise significant improvements in healthcare efficiency, accuracy, and fairness. However, without careful development, these systems might unintentionally exacerbate healthcare disparities. Subtle variations in medical imaging, influenced by factors like a patient's race, gender, or insurance status, may be imperceptible to humans but can be leveraged by AI algorithms, potentially leading to biased predictions. This project seeks to uncover hidden signals in AI models and evaluate their potential effects on fairness. We will utilize three extensive medical imaging datasets from Emory University, encompassing a variety of patient populations and outcomes. Using advanced generative techniques, we will generate synthetic counterparts of these datasets based on demographics, social factors, and clinical variables. Deep learning models will then be trained on both real and synthetic images to predict sociodemographic factors and measure the persistence of hidden signals across data types. By exposing models to manipulated synthetic data, we aim to identify potential unfair associations undetectable in actual patient data. Additionally, we will condense images into mathematical representations and conduct predictive experiments to ascertain if shortcuts are present. Radiologists will analyze simulated videos transitioning images between demographic groups to visualize results and determine if these shortcuts hold clinical relevance. Our objective is to develop tools and guidelines fostering the creation of reliable AI systems in medical imaging to enhance patient outcomes across all groups. Overall, this project combines innovative data generation, predictive modeling, and human-AI collaborative evaluation to proactively identify and mitigate unfair bias in healthcare AI. Our findings will help prevent AI from perpetuating disparities and ensure it fulfills its potential to provide equitable quality care for diverse populations.
