Discharge Prediction Models for Operational Support: The Case of Multisite Healthcare Systems

Presentation Time: 04:45 PM - 05:00 PM

Abstract Keywords: Machine Learning, Healthcare Quality, Data Mining
Primary Track: Applications
Programmatic Theme: Clinical Informatics

Managing multisite health systems operations is a challenging task. The use of predictive analytics has been proposed as an alternative to improve operational and patient outcomes. However, applications of previous work are limited to patient level long term discharge predictions. Such models are useful to estimate patient level outflows but fail to provide support for capacity management decisions that need to be taken in real time, such as bed placement. In this study, we conduct two experiments using more than 140,000 discharge records from two facilities to evaluate single and multitask models to dynamically predict short term discharge volume. In experiment 1, we study the performance of different machine learning models to predict discharges in the next hour and discharges in the next four hours. Additionally, we compare multitask learning models with single task learning models. In experiment 2, we evaluated the performance of a random forest model to predict the number of discharges from 12:00 PM to 4:00 PM with one to four hours in advance. Results from the numerical experiments suggest that a random forest regressor can significantly outperform a linear regression model in most prediction tasks. In addition, we found that predicting discharges in the next hour is harder relative to discharges in the next four hours and that that accurate forecasts of afternoon discharges can be made using a simple set of explanatory variables even when predicting hours in advance.

Speaker(s):
Fernando Acosta-Perez, B.S.
University of Wisconsin-Madison

Author(s):
Fernando Acosta-Perez, B.S. - University of Wisconsin-Madison; Justin Boutilier, Ph.D. - University of Wisonsin-Madison; Gabriel Zayas-Cabán, Ph.D. - University of Wisconsin-Madison; Sabrina Adelaine, Ph.D. - UW-Health; Frank Liao, PhD - University of Wisconsin, Madison - UW Health; Brian Patterson, MD MPH - University of Wisconsin-Madison;