Forecasting Surgical Bed Utilization: Architectural Design of a Machine Learning Pipeline Incorporating Predicted Length of Stay and Surgical Volume.

Abstract

The objective of this study was to develop a machine learning model utilizing data from the electronic health record (EHR) to model length of stay and daily surgical volume, in order to subsequently predict daily surgical inpatient bed utilization. Machine learning is increasingly used to aid healthcare decision-making and resource allocation. Surgical inpatient bed utilization is a key metric of hospital efficiency and an ideal target for optimization. EHR data from all surgical cases over one year at a single institution was obtained. Data from the first 32 weeks of the year were used to train the model with the remaining data used to validate and test the models. Various machine learning approaches were explored to predict hospital length of stay and surgical volume. Seasonal Autoregressive Integrated Moving Average (SARIMA) was used to forecast daily surgical bed requirements. The root mean squared error (RMSE) was reported. For predicting bed utilization > 2 weeks in the future, our optimized models improved prediction from an RMSE of 43.1 to 24.4 beds. For predicting bed utilization in 2 weeks, our optimized models improved prediction from an RMSE of 42.6 to 24.8 beds. Finally, predicting bed utilization same day demonstrated an RMSE of 22.7 beds. We described the architecture of a machine learning approach to forecast surgical bed utilization. Forecasting use of surgical resources may decrease stress on a hospital system through more accurate predicting of the ebbs and flows of hospital needs.