A Decision-Focused Approach to Optimizing Hospital Patient Flow

Abstract

Views of US healthcare have been declining in recent years, with Americans rating their satisfaction with US healthcare quality in 2023 at an all-time low within the past two decades. Given that a majority of healthcare in the US is provided by hospitals, improving the quality of hospital patient care is a natural way to target improving the overall quality of US healthcare. We focus specifically on optimizing hospital-wide patient flow, the movement of patients in a hospital. This thesis takes a novel approach, to the best of our knowledge, to optimize hospital-wide patient flow using decision-focused learning (DFL), also known as a one-stage or end-to-end predict-then-optimize approach. We seek to improve upon current works utilizing two-stage predict-then-optimize approaches, which fail to take into account the impact of predictions on decisions made by the optimization model. We formulate two mixed-integer optimization problems (MIP) to first, make patient bed assignments at the immediate next step and secondly, to do so while accounting for patient demands throughout the rest of the day. The costs in our MIPs, which express future patient care needs, are realistically unknown and predicted using the DFL framework. To train practical models, we use a fictional hospital and training data sourced from the MIMIC-IV database, which contains patient data from the Beth Israel Deaconess Medical Center [19]. We find that using DFL to predict the costs can result in more optimal patient bed assignments in comparison to the two-stage method, consequently reducing decision regret. Additionally, we also find that a hybrid method utilizing both DFL and the two-stage approach performs better than just a two-stage approach and can reduce the error introduced through the two-stage method. We further provide insights on how costs are predicted using DFL to account for this error. Ultimately, we find merits in using DFL to optimize hospital-wide patient flow, and demonstrate the need for future works to further investigate the applications of DFL to healthcare.