Algorithmic fairness audits in intensive care medicine: artificial intelligence for all?

Abstract

[...]a common understanding of protected personal characteristics (e.g., age, gender, and ethnicity) that, at minimum, should be obtained is crucial to adequately design and perform fairness audits. [...]based on the former, relevant protected personal characteristics need to be routinely and uniformly collected in patient health records, worldwide. [...]we need to determine which metrics should be used to assess fairness; are standard AI performance metrics (discrimination and calibration) sufficient or do we need fairness-specific metrics?