JMIR Publications

ABSTRACT

Background:

Emergency department (ED) admissions are one of the most critical decisions made in healthcare, with 40% of admissions originating from the ED. A main challenge with the ED admissions process is the inability to move patients from the ED to an inpatient unit (IU) quickly. Identifying hospital discharge volume in advance may be a useful tool in helping hospitals identify capacity management mechanisms to reduce ED boarding, such as transferring low complexity patients to neighboring hospitals. Although previous research has studied the prediction of discharges in the context of inpatient care, most of the work is on long-term predictions (i.e., discharges within the next 24 to 48 hours) in single site healthcare systems. In this study, we approach the problem of inpatient discharge prediction from a system wide lens and evaluate the potential interactions between the two facilities in our partner multisite system to predict short term discharge volume.

Objective:

To predict discharges from the general care units within a large tertiary teaching hospital network in the Midwest and evaluate the impact of external information from other hospitals on model performance.

Methods:

We conducted two experiments with 174,799 discharge records from two hospitals. In Experiment 1, we predicted the number of discharges across two time points (within the hour and the next four hours) using random forest (RF) and linear regression (LR) models. Models with access to internal hospital data (i.e., system agnostic) were compared with models with access to additional data from the other hospital in the network (i.e., system aware). In Experiment 2, we evaluated the performance of a random forest model to predict afternoon discharges (i.e., 12:00 PM to 4:00 PM) one to four hours in advance.

Results:

In Experiment 1 and Hospital 1, RF and LR models performed equivalently, with R2 scores varying from 0.76 (hourly) to 0.89 (four hours). In Hospital 2, the RF model performed best, with scores varying from 0.68 (hourly) to 0.84 (four hours), while scores for LR models ranged from 0.63 to 0.80. There was no significant difference in performance between a system-aware approach and a system-agnostic one. In experiment 2, the mean absolute percentage error increased from 11% to 16% when predicting four hours in advance relative to zero hours in Hospital 1 and 24% to 35% in Hospital 2.

Conclusions:

Short-term discharges in multisite hospital systems can be locally predicted with high accuracy, even when predicting hours in advance.