Demand Prediction for Better Hospital Capacity Management

Abstract

Accurate hospital bed demand forecasting is critical for ensuring effective patient care and efficient resource allocation. This study evaluates various statistical and machine learning methods to predict daily and hourly inpatient admissions, separations, and emergency department (ED) presentations up to one year in advance. The Advanced Demand Prediction Tool (ADePT) is introduced, which leverages the SARIMAX time series model to capture trends, seasonal patterns, and public holiday effects. Its performance is evaluated using data from a large provider of tertiary health services in Melbourne, Australia against five other statistical and machine learning forecasting models, including rolling window, six-week rolling average, negative binomial regression, an ensemble approach, and random forest regression. The results demonstrated that ADePT generally outperformed other methods when predicting inpatient admissions and separations for multiple forecast horizons. For ED presentations, differences in accuracy were not statistically significant. Importantly, ADePT also showed high accuracy when applied to smaller patient subgroups, including emergency and elective inpatient admissions. By providing reliable short-Term and long-Term forecasts, ADePT could support more effective daily bed management as well as improved long-Term capacity planning.