Predicting medicine inpatient discharges at Massachusetts General Hospital

Abstract

At Massachusetts General Hospital, inpatients often experience significant non-clinical delays in patient care, and frequently wait in the Emergency Department or in inpatient-floor hallways before receiving bed assignments. Such delays result in overcrowding in the Emergency Department, heightened dissatisfaction among patients, and an increase in overall patient length-of-stay. Delays in bed assignments primarily occur because of the discrepancy between the timing of admissions, which generally occur throughout the day, and the timing of discharges, which typically occur in the afternoon. Furthermore, although bed managers know about scheduled admissions in advance, there is no standardized protocol that allows bed managers at the Admitting Department to identify which patients are ready to leave the hospital. In this project, we develop a discharge prediction tool that identifies medicine and neurology inpatient discharges that will occur within the next 24 hours. The goal is to use this tool to enable a more proactive bed-management process at MGH, provide the hospital staff with a methodical way to identify discharges, and ameliorate overcrowding challenges in the Emergency Department. The model was trained using the data of 60,993 inpatients who were hospitalized sometime between May 2016 and September 2018. The prediction algorithm achieved a 0.830 mean AUC-ROC (SD 0.002), 47.6% precision (24 hours), 67.4% precision (48 hours), and 43.8% recall using a decision threshold of 0.31. For inpatients who were on cardiology floors within the Department of Medicine, the model achieved 58.3% precision (24 hours), 74.3% precision (48 hours), and 63.5% recall using 0.31 as the decision threshold. Since the model used data that is accessible in most hospital information systems, it can be applied to other hospitals as well.