Forecasting influenza hospital admissions within English sub-regions using hierarchical generalised additive models

Abstract: Background: Seasonal influenza causes a substantial burden on healthcare services over the winter period when these systems are already under pressure. Policies during the COVID-19 pandemic supressed the transmission of season influenza, making the timing and magnitude of a potential resurgence difficult to predict. Methods: We developed a hierarchical generalised additive model (GAM) for the short-term forecasting of hospital admissions with a positive test for the influenza virus sub-regionally across England. The model incorporates a multi-level structure of spatio-temporal splines, weekly seasonality, and spatial correlation. Using multiple performance metrics including interval score, coverage, bias, and median absolute error, the predictive performance is evaluated for the 2022/23 seasonal wave. Performance is measured against an autoregressive integrated moving average (ARIMA) time series model. Results: The GAM method outperformed the ARIMA model across scoring rules at both high and low-level geographies, and across the different phases of the epidemic wave including the turning point. The performance of the GAM with a 14-day forecast horizon was comparable in error to the ARIMA at 7 days. The performance of the GAM is found to be most sensitive to the flexibility of the smoothing function that measures the national epidemic trend. Interpretation: This study introduces a novel approach to short-term forecasting of hospital admissions with influenza using hierarchical, spatial, and temporal components. The model is data-driven and practical to deploy using information realistically available at time of prediction, addressing key limitations of epidemic forecasting approaches. This model was used across the winter for healthcare operational planning by the UK Health Security Agency and the National Health Service in England.