Research Involving Artificial Intelligence

The AI field is changing fast, and its use in human subjects research can vary widely. For instance, you might use predictive AI to create an algorithm that forecasts the likelihood of a disease or how well someone will respond to a treatment. Alternatively, you might use generative AI for different parts of clinical research, like designing trials or analyzing data.

No matter how you use AI in your research, you need to clearly explain its use in your study plan. This way, the IRB can properly review and approve it.

When you use AI in human subjects research, it usually fits into two main categories:

1. Research focused on AI: Here, your study is about the AI itself. You might be developing or testing the AI. For example, you could be creating an image analysis system that detects cancer in CT or MRI scans, or testing how accurate an algorithm is at diagnosing a disease based on clinical factors and genetic information.
2. Using an existing AI model: In this case, you use an existing AI model as part of your research, but you are not evaluating its safety or effectiveness. Examples include using AI to collect clinical data from medical records, AI tools for image analysis (whether for research or diagnostic purposes), or using AI to assign participants to different groups in a study.

No matter how you use AI in your research, you need to clearly explain its use in your study plan. This helps the IRB properly review and approve your research.

You need IRB review for AI research if it meets the Common Rule or FDA definitions of human subjects research or a clinical investigation. Most researchers know the Common Rule definition, which excludes research using only de-identified data. However, FDA regulations do not make a distinction between identifiable and de-identified data.

So, some AI research using only de-identified data may still need IRB review. For instance, if you use AI and machine learning to analyze human data to detect individuals who might develop a specific disease or condition, you need IRB review, even if all the data is de-identified. This requirement applies even during the model training stage if you are using human data. FDA regulations treat the data itself as a human subject, like biological specimens. If you are unsure whether your research project requires IRB review, we encourage you to seek input from the Office of Human Subjects Research Protections (OHSRP) and the NIH IRB.

When you use software, including AI, in a research study, it might be considered a medical device. If the AI is "intended for use in the diagnosis of disease or other conditions, or in the cure, mitigation, treatment, or prevention of disease in humans or animals" (section 201(h)(1) of the Food, Drug, and Cosmetic Act), it may qualify as a medical device. If the AI is the focus of the investigation, it could be seen as an investigational medical device. In this case, your research must follow FDA regulations and undergo IRB review.

The 21st Century Cures Act excludes certain software functions from being classified as medical devices if their intended use is for:

• Administrative support of a healthcare facility: For instance, software used for billing or patient scheduling.
• Encouraging a healthy lifestyle without referencing any disease or condition: This includes apps for weight management, stress reduction, or relaxation, if they don't claim to diagnose, cure, mitigate, treat, or prevent any disease or condition.
• Electronic patient records.
• Transferring, storing, converting formats, or displaying clinical laboratory test or other device data, unless the software is intended to interpret or analyze the data.
• Clinical decision support software (CDSS), provided it meets certain requirements.
• For Clinical Decision Support Software (CDSS) to be excluded from the definition of a medical device, it must meet all the following criteria:
• It is not intended to acquire, process, or analyze a medical image, a signal from an in vitro diagnostic device, or a pattern or signal from a signal acquisition system.
• It is intended to display, analyze, or print medical information about a patient or other medical information, such as peer-reviewed clinical studies and clinical practice guidelines.
• It is intended to support or provide recommendations to a healthcare professional about the prevention, diagnosis, or treatment of a disease or condition.
• It is intended to enable a healthcare professional to independently review the basis for the recommendations** provided by the software, ensuring that healthcare professionals do not rely primarily on the software's recommendations to make clinical diagnoses or treatment decisions for individual patients.

These determinations can be complex and difficult. The FDA offers an online Digital Health Policy Navigator to help you determine if your software, whether AI or otherwise, might be considered a medical device. If you are unsure whether your AI qualifies as a medical device, we encourage you to contact the Office of Human Subjects Research Protections (OHSRP) for guidance.