Where do doctors disagree? Characterizing Decision Points for Safe Reinforcement Learning in Choosing Vasopressor Treatment

In clinical settings, domain experts sometimes disagree on optimal treatment actions. These "decision points" must be comprehensively characterized, as they offer opportunities for Artificial Intelligence (AI) to provide statistically informed recommendations. To address this, we introduce a pipeline to investigate "decision regions", clusters of decision points, by training classifiers for prediction and applying clustering techniques to the classifier's embedding space. Our methodology includes: a robustness analysis confirming the topological stability of decision regions across diverse design parameters; an empirical study using the MIMIC-III database, focusing on the binary decision to administer vasopressors to hypotensive patients in the ICU; and an expert-validated summary of the decision regions' statistical attributes with novel clinical interpretations. We demonstrate that the topology of these decision regions remains stable across various design choices, reinforcing the reliability of our findings and generalizability of our approach. We encourage future work to extend this approach to other medical datasets.