Key points are not available for this paper at this time.
Introduction The rapid expansion of artificial intelligence (AI) has produced many opportunities, but also new risks that must be actively managed, particularly in the health care sector with clinical practice to avoid unintended health, economic, and social consequences. Methods Given that much of the research and development (RD) involving human subjects is reviewed and rigorously monitored by institutional review boards (IRBs), we argue that supplemental questions added to the IRB process is an efficient risk mitigation technique available for immediate use. To facilitate this, we introduce AI supplemental questions that provide a feasible, low-disruption mechanism for IRBs to elicit information necessary to inform the review of AI proposals. These questions will also be relevant to review of research using AI that is exempt from the requirement of IRB review. We pilot the questions within the Department of Veterans Affairs–the nation's largest integrated healthcare system–and demonstrate its efficacy in risk mitigation through providing vital information in a way accessible to non-AI subject matter experts responsible for reviewing IRB proposals. We provide these questions for other organizations to adapt to fit their needs and are further developing these questions into an AI IRB module with an extended application, review checklist, informed consent, and other informational materials. Results We find that the supplemental AI IRB module further streamlines and expedites the review of IRB projects. We also find that the module has a positive effect on reviewers' attitudes and ease of assessing the potential alignment and risks associated with proposed projects. Discussion As projects increasingly contain an AI component, streamlining their review and assessment is important to avoid posing too large of a burden on IRBs in their review of submissions. In addition, establishing a minimum standard that submissions must adhere to will help ensure that all projects are at least aware of potential risks unique to AI and dialogue with their local IRBs over them. Further work is needed to apply these concepts to other non-IRB pathways, like quality improvement projects.
Makridis et al. (Thu,) studied this question.