Operationalizing Reparative Algorithmic Impact Assessments for Healthcare AI Governance

Algorithmic impact assessments (AIAs) and fairness audits, as currently practiced, can miss dimensions that shape whether AI deployments benefit or harm the populations they target. In healthcare, particularly across low-resource settings in the Global Majority, these include power asymmetries between developers, institutions, and affected communities; the political economy and sovereignty conditions of training data; the epistemic exclusion of non-Western knowledge systems; and the distribution of technical capacity between deployment partners as systems persist over time. This paper specifies a Reparative Algorithmic Impact Assessment (R-AIA) protocol that augments standard AIA workflows along these dimensions. Building on prior work, the protocol combines six interconnected assessment steps with four auditable artifacts: provenance audit trails, co-constructed impact taxonomies, community-defined evaluation rubrics, and capacity-transfer metrics. We provide structured specifications for each artifact, including a relational schema and audit query for provenance, and illustrate how the protocol changes the assessment of a hypothetical maternal-health deployment. We also map the artifacts onto existing governance frameworks and consider implementation burden, capture, interoperability, and empirical validation. Taken together, we strive to offer a minimum viable instantiation for resource-constrained settings. This R-AIA protocol is intended to complement rather than replace conventional fairness, safety, privacy, and performance assessment by making additional structural conditions inspectable in practice. And while the paper develops the protocol for healthcare AI, we anticipate that its underlying assessment logic may prove useful in other high-stakes deployment settings.