Predicting surgical inpatients' discharges at Massachusetts General Hospital

Abstract

In the last few years, MGH has grappled with severe bed capacity management problems. As a result, delays occur in delivering the patient to the right bed at the right time, hindering patient care. One of the root causes for those delays is the mismatch between the timing of admissions and discharges. Particularly, while bed managers know about most admissions well in advance, there is a prevalent lack of central transparency regarding which patients might be ready to leave the hospital and what are the barriers that may delay their discharge. This project aims to improve MGHs bed management processes by introducing a predictive model (based on neural network) that identifies, in real time, surgical inpatients discharges that will occur in the next 24 hours. As part of this research, we present a new modeling methodology, formalizing concepts of 'Milestones to Post-Operative Recovery' and 'Barriers to Discharge', which systematically track patients progress towards discharge. For every admitted surgical patient, our solution outputs a score that is correlated with the likelihood for discharge within 24 hours, and derives a list of barriers to discharge ranked by their significance. In addition, the solution predicts with high accuracy (R-Square 0.86) the total number of daily surgical inpatient discharges, a key piece of information for bed managers. Given training population of 15,553 surgical inpatients admitted to MGH between May 2016 and August 2017, and test population (out-of-sample) of 1,151 surgical inpatients hospitalized during September 2017, the model achieved remarkable performance with ROC of 0.857. During non-holiday weekdays, among the top 10 ranked surgical inpatients identified by the algorithm to have the highest probability of being discharged, 90% were discharged within 24 hours and 97% were discharged within 48 hours, capturing 23% of the hospital's daily surgical discharges. Among the top 30 patients ranked by the algorithm, 69% were discharged within 24 hours and 89% were discharged within 48 hours, capturing 53% of the hospital's daily surgical discharges. The model was implemented as a web-based tool and is currently being piloted at MGH. Preliminary results show potential to promote proactive discharge processes to eliminate unnecessary delays. The implemented solution is using standard EMR data streams, and can be generalized across hospitals.