Presenting on “Varieties of Uncertainty in AI- Assisted Medicine”.

Anya Plutynski, PhD, professor of philosophy at Washington University in St. Louis, will speak on Thursday, October 17, 2024 at 10:00 am CT in Whitaker 218.

Registration is required to attend virtually. Please register here.

Abstract: Advances in technology can disrupt our conceptual frameworks, and as a result, clinical practices. Indeed, they can lead to the invention of entirely new categories. For instance, advances in screening for cancer have led to the identification of “precancers,” or “indolent lesions of epithelial origin,” (IDLE), a label created to name those lesions that are unlikely to cause harm if they are left untreated (Esserman, et. al., 2014).

In this talk, I consider the ways in which AI has led to a transformation in the ways “uncertainty” arises in medical decision making. The literature on medical decision making historically located uncertainty in an agent – a researcher, policymaker, clinician, or patient. An agent might be uncertain about the disease classifications that they are relying upon, the quality or clinical relevance of the evidence they are relying upon, the reliability of an instrument, effectiveness of a treatment, or a patients’ values or goals. All these uncertainties in turn shape the complex decision-making processes involved in diagnosis, prognosis, or treatment.

In contrast, when engineers and computer scientists speak of quantifying the “uncertainty” of AI, this is a property of the prediction, which is, in turn, a function of the quality or comprehensiveness of either the training data or the model or algorithm. “Uncertainty” in discourse around AI is thus not a property of agents, per se. Or perhaps better, it is only secondarily a property of agents that use the AI.

As these two communities are increasingly coming into conversation with one another, researchers and clinicians need to translate talk about various measures or types of uncertainty from one context to another. Failure of communication or clarity can lead to harm. So, I offer a taxonomy of different senses of “uncertainty” at play in this context, using case studies in AI into clinical practice, in service of greater clarity.