Table 2.

Strategies to address AI^a equity issues that were abstracted from the literature.

Category and strategy			Description
Background Context
	Foster diversity	Create AI development teams with diverse characteristics, experiences, and roles to increase consideration of equity throughout development and decrease blind spots.
	Train developers and users	Train AI developers and users in equity considerations and the ethical implications of AI, as these topics may be unfamiliar to some.
	Engage the broader community	Foster community involvement throughout development, from conception to postdeployment, to increase the likelihood that developers prioritize equity concerns.
	Improve governance	Enact robust regulation and industry standards to align AI applications with social norms, including equity, safety, and transparency.
Data Characteristics
	Improve diversity, quality, or quantity of data	Train models with large, diverse samples that are representative of the target population for the application and contain all relevant features.
	Exclude sensitive variables to correct for bias	Exclude sensitive variables or replace them with variables that are more directly relevant to health outcomes to prevent models from discriminating directly on these characteristics.
	Include sensitive variables to correct for bias	Include sensitive variables to improve model accuracy, increase explanatory power, and enable easier testing for inequitable impact.
Model Design
	Enforce fairness goals	Formulate a fairness norm and enforce it in the model by editing the input data, objective function, or model outputs.
	Improve interpretability or explainability of the algorithm	Choose models that are inherently explainable (such as decision trees), build models with post hoc explainability, or explore explainable local approximations to model decision making.
	Evaluate disparities in model performance	Evaluate model performance on a wide range of metrics across subgroups, particularly groups that might face inequitable impact, then report and act upon the results.
	Use equity-focused checklists, guidelines, and similar tools	Incorporate equity-focused checklists into workflows for developers, reviewers of AI models, health care providers using an application, or patients who want to understand algorithm outputs.
Deployment Practices
	Increase model reporting and transparency	Provide more information on AI equity issues, including publishing standardized equity-related analyses on models, increasing independent model reviews, and requiring equity discussions in academic journals.
	Seek or provide restitution for those negatively impacted by AI	Proactively provide restitution to those harmed by AI or create legal frameworks so they can seek restitution.
	Avoid or reduce use of AI	Consider discontinuing model use if equity sequelae are severe or if improvement efforts have been fruitless.
	Provide resources to those with less access to AI	Improve access to AI for disadvantaged groups and low-income countries by subsidizing infrastructure, creating education programs, or hosting AI conferences in these locations.

^aAI: artificial intelligence.