Artificial Intelligence in the COVID-19 Response

This study was prepared for the Patient-Centered Outcomes Research Institute (PCORI) Emerging Technology and Therapeutics Report series. This study, on strategies to improve the impact of artificial intelligence (AI) on health equity, is the second of 2 reports examining AI in health care.

Problem Statement

The use of AI in clinical care, public health, and health system administration has expanded rapidly in recent years. AI applications in health care have the potential to improve accuracy, personalization, and fairness, but may also introduce new biases or perpetuate existing inequalities as a result of data limitations and other challenges. At the same time, the COVID-19 pandemic has underscored the persistence of health disparities in the United States and abroad, with disadvantaged populations facing high rates of infection, hospitalization, and death. The pandemic has provided further evidence that AI developers, users, and policy makers will increasingly need strategies to mitigate negative impacts and enhance positive impacts of AI on health equity.

Approach

We conducted a scoping review to identify strategies used to address equity issues posed by AI in health care. We conducted interviews with stakeholders to inform our research questions and guide our study design. We searched, screened, and reviewed a wide range of academic and gray literature as part of our study.

Key Findings

• We identified 18 equity-related issues that are raised by the use of AI in health care. Concerns about unrepresentative and biased data are most commonly mentioned in the literature. Other prominent issues include balancing any potential trade-offs between model accuracy and fairness, biased or nonrepresentative AI developers, and limited information on population characteristics.
• We identified 15 strategies proposed to address these equity-related issues posed by the use of AI in health care. The most commonly proposed strategies in the literature were evaluating disparities in model performance, improving data inputs, engaging the broader community in AI development, and improving governance. Most issues are complex and are likely best addressed through multiple complementary strategies.
• In some cases, AI can be used to address long-standing problems of health inequity, whose causes are rooted in societal issues that are independent of the use of AI in health care. This includes implementing AI models that are less biased than current decision-making practices, using AI to better understand the extent and cause of health disparities, using AI to target health service delivery to those who need it most, and fielding AI that directly improves marginalized communities' access to care.
• Efforts to improve the impact of AI on health equity could benefit from further research on which strategies have proved the most effective in real-world settings as well as on best practices for strategy implementation.