Near-term forecasting of Covid-19 cases and hospitalisations in Aotearoa New Zealand

Near-term forecasting of infectious disease incidence and consequent demand for acute healthcare services can support capacity planning and public health responses. Despite well-developed scenario modelling to support the Covid-19 response, Aotearoa New Zealand lacks advanced infectious disease forecasting capacity. We develop a model using Aotearoa New Zealand’s unique Covid-19 data streams to predict reported Covid-19 cases, hospital admissions and hospital occupancy. The method combines a semi-mechanistic model for disease transmission to predict cases with Gaussian process regression models to predict the fraction of reported cases that will require hospital treatment. We evaluate forecast performance against out-of-sample data over the period from 2 October 2022 to 23 July 2023. Our results show that forecast performance is reasonably good over a 1-3 week time horizon, although generally deteriorates as the time horizon is lengthened. The model has been operationalised to provide weekly national and regional forecasts in real-time. This study is an important step towards development of more sophisticated situational awareness and infectious disease forecasting tools in Aotearoa New Zealand. Author summary The emergency phase of the Covid-19 pandemic has ended, but Covid-19 continues to put significant additional load on stretched healthcare systems. Forecasting the number of hospital cases caused an infectious disease like Covid-19 over the next few weeks can help with effective planning and response. The ability to forecast reliably requires timely, high-quality data and accurate mathematical models. We have developed a model for forecasting the number of Covid-19 cases and hospitalisations in Aotearoa New Zealand. The model works in two stages: firstly predicting the number of new cases and secondly estimating the proportion of those cases that will need hospital treatment. The model produces a range of likely values, which is important because is impossible to predict with 100% accuracy. We show that the model does a reasonably good job of predicting hospitalisations up to 3 weeks ahead. The model has been used by public health agencies in Aotearoa New Zealand to help with healthcare capacity planning.