Mert R. Sabuncu, Ph.D., Alan Q. Wang, Ph.D., and Minh Nguyen, M.S.
Artificial intelligence (AI) promises to be a transformative technology for medicine and health care. As such, there is an increasing interest in ensuring its ethical use. In this perspective, we consider the employment of AI for medical diagnostics, where the goal is the detection and classification of an underlying pathology, based on data such as patient information, clinical presentation, tests, and imaging. We argue that instead of prioritizing fairness criteria that measure disparities between protected groups, the primary goal should be to assess and enhance diagnostic accuracy within each subpopulation. This approach shifts the focus from optimizing overall population accuracy to ensuring maximal accuracy in each subpopulation. Our perspective implies that we should be using all available information, including protected group identity, in our methods. We decouple the goal of accurate diagnosis from fairness considerations in screening and postdiagnosis clinical decisions, which often require the allocation of finite resources. Furthermore, we underscore the importance of collecting high-quality and representative datasets for each subpopulation, including ensuring that the ground truth labels we train and with which we evaluate are unbiased.