Comparing statistical forecasting methods for modelling respiratory virus healthcare demand

Abstract Background Emergency departments in the United Kingdom are under sustained pressure with rising demand for healthcare, gradually decreasing the functionality of the NHS. Forecasts can aid management in the allocation of appropriate healthcare supplies and staffing needs in advance of an expected surge, as well as allow stakeholders to assess most likely outcomes and potential worst-case scenarios. This is especially valuable during the seasonal waves of respiratory infections. Respiratory viruses are usually concentrated in a 6-8 week seasonal surge during the winter period, creating large pressure on the health system. Current forecast models’ performance depreciates significantly after 1 week. This analysis will forecast 6 weeks of respiratory health admissions, comparing traditional and newly emerging forecasting methods. Methods Using 20 years of English hospital admissions data, we evaluated both within season training models (2-20 weeks) and longer term seasonal models (1-10 years), to forecast 6 weeks of hospital admissions. Respiratory admissions were defined according to ARI/ILI ICD10 codes assigned in hospital, stratified by age and region. Models were iteratively constructed, trained and probabilistic forecasts were evaluated across the respiratory season of September to April using Weighted Interval Score. Averages of each model across iterations were generated for comparison. Results For non-seasonal models, shorter model training lengths created better forecast models for 6 weeks. For seasonal models, optimal training length varied across model complexity. Simple models performed best at forecasting 1 week, however performance deteriorated the most as the forecast window increased. Overall, the Prophet model performed best on average at forecasting 6 weeks across all scenarios. Conclusions Applying our framework for optimal model selection in forecasting analysis will help to inform better practice for future health demand modelling. Key messages • This analysis will inform better practice for future health demand forecasting 6 weeks in advance, allowing for preparation of resources for seasonal surges in healthcare demand. • This analysis suggests using models with modest levels of increased complexity over simpler models could significantly improve forecast performance in the short-medium term.