Development of Multiservice Machine Learning Models to Predict Postsurgical Length of Stay and Discharge Disposition at the Time of Case Posting.

Abstract

Objective: Develop machine learning (ML) models to predict postsurgical length of stay (LOS) and discharge disposition (DD) for multiple services with only the data available at the time of case posting.

Background: Surgeries are scheduled largely based on operating room resource availability with little attention to downstream resource availability such as inpatient bed availability and the care needs after hospitalization. Predicting postsurgical LOS and DD at the time of case posting could support resource allocation and earlier discharge planning.

Methods: This retrospective study included 63,574 adult patients undergoing elective inpatient surgery at a large academic health system. We used surgical case data available at the time of case posting and created gradient-boosting decision tree classification models to predict LOS as short (≤1 day), medium (2-4 days), and prolonged stays (≥5 days) and DD as home versus nonhome.

Results: The LOS model achieved an area under the receiver operating characteristic curve (AUC) of 0.81. Adding relative value unit and historical LOS through the similarity cascade increased the accuracy of short and prolonged LOS prediction by 9.0% and 3.9% to 72.9% and 74%, respectively, compared with a model without these features (P = 0.001). The DD model had an AUC of 0.88 for home versus nonhome prediction.

Conclusions: We developed ML models to predict, at the time of case posting, the postsurgical LOS and DD for adult elective inpatient cases across multiple services. These models could support case scheduling, resource allocation, optimal bed utilization, earlier discharge planning, and preventing case cancelation due to bed unavailability.