20 - Artificial Intelligence (AI) Adventures in Human Subjects Research Video Series

Background: : Potential risks to participants increase when Artificial Intelligence (AI) is utilized in human subjects research (HSR). Use of AI in human research is growing faster than guidance on how to avoid increased risks. A foundational understanding of AI data can aid in mitigating the risk of using AI in HSR.

Methods: : Content analysis of federal documents and academic journal articles informed the decision to focus specifically on the risks that arise when researchers input data into generative AI. Based on our analysis, a video series titled "AI Adventures in Human Subjects Research" was created. A primary and secondary research approach was used to create the video scripts.

Conclusion: : Development of an overarching concept called “data ancestry” allowed us to categorize the different types of previously collected and stored research data that researchers feed into generative AI models. We categorized older secondary research data that may be historically biased as Grandfather Data. Aunt Data is categorized as de-identified publicly available data sets that can be combined with other auxiliary information. Sister Data is the category for social media content that is publicly available on the internet and false or fabricated data is categorized as Uncle Data.

With the different types of secondary research data classified, we were able to draw out the risks that emerge when researchers input these data ancestors in AI models. For example, we were able to illustrate how de-identified data can be re-identified when an Aunt Data set is put into an AI model along with other data sets.

Once the risks were more clearly defined, we were able to develop strategies for mitigating the potential harm to research participants. For example, researchers that are combining Aunt Data Sets can remove, generalize, or mask data variables to make it more difficult for an AI model to match data points and re-identify individuals. We concluded the risk mitigation strategies should be more readily available to researchers. So, a video series was created to summarize concepts of data ancestry and AI data input strategies.

Limitations:: AI Adventures in Human Subjects Research videos are currently limited on two fronts. First, they focus on understanding the risk of inputting data into AI models. Risks from data generated by AI models need to be investigated. Second, the series addresses general strategies. One could argue that there is a need to develop approaches for using AI in medical research involving human subjects that extends past the initial inputting of data in AI models.

Discussion: : To overcome the limitations of the video series, we will investigate the potential risks of AI generated data in research involving human subjects. Even if researchers employ AI data input strategies, AI-generated data may create additional risks that require further mitigation.

Another plan is to create more AI Adventures in Human Subjects Research videos. New videos in the series will help researchers gain an understanding of the regulatory requirements and risks that emerge when AI models are used in all stages of medical research that involves human participants. We will share strategies and safeguards for mitigating those risks.