Considerations for IRB Review: Artificial Intelligence & Machine Learning
While artificial intelligence and machine learning are not new concepts in science, in the last decade the clinical research community has seen tremendous growth in these areas. Researchers are developing new applications at a rapid pace, and while it is exciting, it is also maybe a bit scary at the same time.
But what do “artificial intelligence” and “machine learning” really mean?
- Artificial intelligence (AI): A feature where machines learn to perform tasks, rather than simply carrying out computations that are input by human users. (NIH)
- Machine learning: An approach to AI in which a computer algorithm is developed to analyze and make predictions from data that is fed into the system. (NIH)
- Algorithm: An overarching term describing any processes or set of rules to be followed in calculations or other problem-solving operations. An algorithm takes some input and uses mathematics and logic to produce the output. Broadly, it refers to a step-by-step procedure for solving a problem. (Merriam-Webster)
Technically, “if/then” is an algorithm. Statistical tests are also algorithms, but they are different than AI algorithms, which are a continual process of providing new inputs.
AI algorithms take both inputs and outputs simultaneously in order to “learn” the data and produce appropriate outputs when given new/unfamiliar inputs. AI algorithms typically consist of a collection of algorithms. In machine learning, AI algorithms are “fed” data and are asked to process it. An AI system can make assumptions, test, and learn autonomously.
Those AI “assumptions” have to be managed or confined; in other words, there must be some human control over the assumptions. In 2017, Facebook Artificial Intelligence Research (FAIR) trained chatbots on a collection of text conversations in English between humans playing a simple trading game involving balls, hats, and books. When programmed to experiment with the English language and tasked with optimizing trades, the chatbots seemed to evolve a reworked version of English to better solve their task.
To quote recent thoughts on ethics and machine learning, “Academia and wider society have laid down ethical principles as a way to ward off a repeat of bitter historical events, but it certainly seems that these will be eroded by uncertainties about consent, harm, and even what constitutes a human subject.” Such statements suggest a need for more specific guidelines in this space, which are currently scarce.
The Department of Defense (DoD) has developed guidance suggesting research involving AI should be designed to be:
With that guidance in mind, institutional review boards (IRBs) may want to include the following additional considerations when reviewing research involving AI and machine learning:
- Remember the regulatory concepts of “identifiable” and “private” information. Navigating these provisions can be complex when you consider the large amounts of human data needed for AI development.
- Preserve informed consent to use private, personal identifying data as much as possible.
- Informed consent is important when using private or personal identifying information.
- Review the potential for bias in the algorithms used for “learning” and the “assumptions” used to learn, such as:
- Obtain scientific review to assess potential bias
- Avoid algorithmic bias by documenting and defining how the data is selected
- Show how bias is mitigated
- Exempt research (such as use of data) must still comply with the Belmont Report.
Recently, the DoD changed the provision for scientific review of research to include minimal risk research. It could very well be because of the increasing volume of AI research, as more research into AI is developed. While this idea seems scary, IRBs are well equipped to handle ethical review and oversight of artificial intelligence research: after all, it starts with human data.