There is widespread agreement that fairness—ensuring equitable performance across similarly situated individuals and across groups—is a fundamental principle for ethical development of artificial intelligence for health care (AI-HC). In the U.S., the NIH has launched initiatives to address unfairness due to lack of diversity in datasets, such as the All of Us Research Program and the Human Pangenome Reference Consortium. Other efforts focus on diversity and representation among researchers. However, it is not yet clear whether these initiatives are sufficient to achieve fairness in AI-HC. Concepts and practices associated with diversity and representation shape the extent to which AI-HC researchers and developers achieve goals for fairness. Yet insufficient understanding of how diversity and representation are conceptualized and put into practice within AI-HC projects is an impediment to achieving fairness in AI-HC datasets. Prior studies of biomedical research indicate that scientists often hold differing concepts of diversity (e.g. genetic or other biomarkers, self- reported race/ethnicity) and representation (e.g., the inclusion of specific historically underrepresented groups in datasets vs. addressing how structural inequalities impact the data for underrepresented groups). In turn, these concepts shape implementation of the practices used to achieve diversity and representation in datasets, such as diversification of researchers, research participants, or technical solutions to bias. We propose to assess how diversity and representation are conceptualized and put into practice in 50 NIH-funded AI-HC research projects. Our analysis will be guided by Steven Epstein’s “inclusion-and-difference paradigm” as a conceptual framework. We will employ a “microethics” perspective, which focuses analysis on how high-level ethical goals, like fairness, are understood and put into practice in technical fields. This perspective allows examination of how different actors (e.g. data scientists, clinicians, annotators) involved in AI dataset development perceive three sets of issues: how diversity and representation are defined; trade-offs in scientific and diversity goals; and the downstream impact on fairness. Informed by these findings, we will develop evidence-informed practical guidance to support the future creation of fair datasets in AI-HC. Our aims are to: Aim 1: Assess data scientists’ concepts and practices relevant to diversity and representation in creating datasets for AI-HC will be achieved through a review of policy and guidance documents (Aim 1a) and interviews of approximately 100 investigators from 50 systematically sampled NIH-funded AI-HC research projects (Aim 1b). Aim 2: Formulate and disseminate practical, evidence-informed guidance to support fairness in AI-HC development will be achieved through a modified Delphi process engaging experts in areas relevant to fairness, diversity and representation in AI-HC. We will use a multi-pronge...