Short term forecast of new daily pandemic hospitalizations: A time series model for a single hospital

Abstract

Reliable hospital admission can aid contingency planning during pandemics. While some studies have developed models for predicting new hospitalizations, most focus on data at the regional or national level. During health crises, a prediction model tailored to a single hospital is essential, as models based on regional data may fail to account for the heterogeneity within the local population in terms of demographics, morbidity, and other relevant characteristics. Our study addresses this gap by presenting an approach for short-term forecasting of daily pandemic admissions to a single hospital. We develop a time series model using COVID-19 admission data during the Alpha wave to one Norwegian hospital with a highly heterogeneous catchment area that includes both urban and rural areas. Previous hospitalizations are included as predictors, along with the local reproduction number (ℜ-number) to capture pandemic trends. To account for demographic differences in admission rates, we employ group-based modelling to divide the catchment area into sub-areas. Forecasts generated from sub-area models are then merged and compared with the forecasts from a model for the entire catchment area. The model's forecasting ability is tested on the Delta wave. The merged model outperforms the total model on the Alpha wave, and both surpass the ARIMA benchmark. On the out of sample Delta wave, the total model performs better overall. While the model overpredicts admissions at the beginning of the Delta wave and the prediction intervals are somewhat conservative, it demonstrates potential for reliably forecasting new daily pandemic admissions. Continuous model adaption will however be necessary as the pandemic evolves.

Keywords: COVID; Forecasting; Hospitalization; Pandemic preparedness; Time-series model.