Table 2.

Guidelines: general premises and recommendations for the inclusion of race in clinical prediction models

Item	Statement
4	There is not a universally consistent approach to conceptualizing, measuring and classifying an individual’s race or ethnicity, although the ‘gold standard’ is typically self-report. (P6)
5	Race or ethnicity should be assessed and defined similarly for model building and application of models in practice, using standards that facilitate consistency (such as the OMB/NIH Standards for the Classification of Federal Data on Race and Ethnicity). a(R1) Modelers should report clearly how race was obtained and defined in their sample. (R2)
6	Patients should be informed by clinicians/health systems when models including race, are used in clinical or resource allocation decisions. E.g., “This prediction makes use of demographic information, such as your age, sex and race, and clinical information, such as…” (R3)
7	Decisions supported by polar and non-polar predictions have different ethical considerations. Polar predictions most frequently arise when models are used for allocation of scarce health resources. (P7; see also P9)
8	Great caution must be exercised when attempting to adapt or use a model for a different clinical decision than the original application, or in a markedly different population. Transportability of the model must be carefully examined, both for bias (see Tables 3, 4) and for fairness (see Table 5) concerns. (R4)
9	When race is included as a candidate variable, model developers must be transparent about the reasoning and: explain the rationale, clearly outlining potential harms (Box 3) and benefits (Box 4), including references to existing models and other relevant prior literature. (R5)

P denotes premise, R recommendation.

^aOMB has recently revised these standards to include a category for “Middle Eastern or North African” (MENA)⁷⁴.