Anesthesiologist Surgery Assignments using Policy Learning

Andrew Ward, Zhengyuan Zhou, Nicholas Bambos, Ellen Wang, David Scheinker

Abstract: Anesthesiologists are currently assigned to surgeries based primarily on anesthesiologist availability and specialty, but optimizing anesthesia time is not generally considered. If certain anesthesiologists perform faster on different patient or surgical cohorts, then incorporating patient-specific and surgery-specific features in scheduling decisions could reduce anesthesia time, and therefore improve operating room efficiency. We formulate the problem of assigning anesthesiologists to surgeries as a policy learning problem. We use random forests and generalized random forests to derive counterfactual estimates, and find the optimal decision tree based on these estimates. We formulate and solve the optimal decision tree problem as a mixed-integer program, and evaluate our decision tree policies using doubly robust estimation techniques. We also demonstrate how our methods can be used to solve a budget-constrained assignment problem by assigning individual costs to each anesthesiologist. The derived policies offer performance improvements over historical scheduling, but are unable to offer larger anesthesia time reductions due to anesthesiologist performance being primarily correlated with prior performance.