A Hierarchical Bayesian Model for Improving Short-Term Forecasting of Hospital Demand by Including Meteorological Information

Summary The effect of weather on health has been widely researched, and the ability to forecast meteorological events can offer valuable insights into the effect on public health services. In addition, better predictions of hospital demand that are more sensitive to fluctuations in weather can allow hospital administrators to optimize resource allocation and service delivery. Using historical hospital admission data and several seasonal and meteorological variables for a site near the hospital, the paper develops a novel Bayesian model for short-term prediction of the numbers of admissions categorized by several factors such as age group and sex. The model proposed is extended by incorporating the inherent uncertainty in the meteorological forecasts into the predictions for the number of admissions. The methods are illustrated with admissions data obtained from two moderately large hospital trusts in Cardiff and Southampton, in the UK, each admitting about 30000–50000 non-elective patients every year. The Bayesian model, computed by using Markov chain Monte Carlo methods, is shown to produce more accurate predictions of the number of hospital admissions than those obtained by using a 6-week moving average method which is similar to that widely used by hospital managers. The gains are shown to be substantial during periods of rapid temperature changes, typically during the onset of cold and highly variable winter weather.