A case-individual data-driven optimization approach for surgery planning

Abstract Operating rooms generate more than 40% of a hospital’s total revenue, making effective surgery planning a highly relevant problem for many hospitals. Surgery planning is a challenging task that requires a decision maker to solve multiple interrelated (sub-)problems under uncertainty. An important source of uncertainty are surgery durations—their realizations can vary substantially, depending on a number of factors related to the type of surgery, as well as patients’ and surgeons’ characteristics. This paper proposes a data-driven approach to solve a complex and practically relevant surgery planning problem using case-individual information contained in electronic health records of past surgeries. Our approach combines machine learning models (Quantile Regression Forests) for estimating durations of individual surgeries and data-driven optimization techniques to assign a set of planned surgeries to operating rooms, and to simultaneously determine their sequence and starting times for a given workday. Numerical analyses based on a real-world dataset of a large university hospital demonstrate the value of our approach. Our approach leads to a cost improvement of 35% compared to a naive approach that reflects the current practice of our case hospital, and we can show that the largest part (approximately 23% points) of this improvement can be attributed to the sequencing and room assignment extensions in our model. We are able to quantify the value of case-individual information, identify the most relevant co-variates (“features”), and demonstrate how better uncertainty estimations affect the individual planning tasks, i.e., room assignment, sequencing, and appointment scheduling.