Overview

• Researchers detailed the method in the Journal of Medical Internet Research on September 4, outlining how it surfaces subgroup biases before model training.
• Evaluations on medical images, patient records, and the National Health and Nutrition Examination Survey using varied model classes uncovered both known and previously overlooked disparities.
• The workflow examines input data and model outputs, including lab values, imaging, predicted diagnoses, and risk scores.
• The team reports adaptability across simple and advanced algorithms, including systems that power large language models, and usability on both small and complex datasets.
• Authors urge use by developers, researchers, and regulators, while cautioning that equitable AI also depends on improved data collection, interpretation, and broader system reforms.