A modular approach to forecasting COVID-19 hospital bed occupancy.

Abstract

Background: Monitoring the number of COVID-19 patients in hospital beds was a critical component of Australia's real-time surveillance strategy for the disease. From 2021 to 2023, we produced short-term forecasts of bed occupancy to support public health decision-making.

Methods: We present a model for forecasting the number of ward and intensive care unit (ICU) beds occupied by COVID-19 cases. The model simulates the stochastic progression of COVID-19 patients through the hospital system and is fit to reported occupancy counts using an approximate Bayesian method. We do not directly model infection dynamics-instead, taking independently produced forecasts of case incidence as an input-enabling the independent development of our model from that of the underlying case forecast(s).

Results: Here, we evaluate the performance of 21-day forecasts of ward and ICU occupancy across Australia's eight states and territories produced across the period March and September 2022. We find forecasts are on average biased downwards immediately prior to epidemic peaks and biased upwards post-peak. Forecast performance is best in jurisdictions with the largest population sizes.

Conclusions: Our forecasts of COVID-19 hospital burden were reported weekly to national decision-making committees to support Australia's public health response. Our modular approach for forecasting clinical burden is found to enable both the independent development of our model from that of the underlying case forecast(s) and the performance benefits of an ensemble case forecast to be leveraged by our occupancy forecasts.