INFORMS Seminar: Machine Learning to Address Hospital Capacity

Want to know how to combine Machine Learning and Optimization in healthcare? Join our seminar with MIT post-doc Dr. Taghi Khaniyev on September 26th, 5:00 – 6:00 PM at Curry 333. To join us, please RSVP here.

Abstract:
Hospital congestion is caused by, among other factors, an imbalance of the timing of patient admissions and discharges. We implemented a real-time clinical support tool to predict next-day’s surgical patient discharges. We trained a machine learning model using 20,745 previous patients’ recovery care paths at a Massachusetts General Hospital (MGH). Then, with an iterative process, we analyzed mismatches between predicted and actual discharges to identify areas for improvement in the discharge process and to enhance the accuracy of the tool. Three pilot implementations were conducted on the selected inpatient floors in collaboration with nurse directors and attending surgeons. The model achieved an average out-of-sample AUC of 0.89 and identified an estimated 128 savable bed-days during the 90-day study period. To enable timely discharges, we also developed an optimization model which identifies a minimal list of barriers that need to be addressed to successfully transition each patient out of hospital. This list would ideally serve as an automatic draft ‘action plan’ for each patient which is of utmost importance, especially at times of capacity crises.