Machine learning forecasts for seasonal epidemic peaks: lessons learnt from an atypical respiratory syncytial virus season

Seasonal peaks in infectious disease incidence put pressures on health services. Therefore, early warning of the timing and magnitude of peak activity during seasonal epidemics can provide information for public health practitioners to take appropriate action. Whilst many infectious diseases have predictable seasonality, newly emerging diseases and the impact of public health interventions can result in unprecedented seasonal activity. We propose a machine learning process for generating short-term forecasts, where models are selected based on their ability to correctly forecast peaks in activity and can be useful during the aforementioned atypical seasonal activity, in contrast to traditional modelling. We have validated our forecasts using typical and atypical seasonal activity, using respiratory syncytial virus (RSV) activity during 2019-2021 as an example. During the winter of 2020/21 the usual winter peak in RSV activity in England did not occur but was deferred until the Spring of 2021. We compare a range of machine learning regression models, with alternate models including different independent variables, e.g. with or without seasonality or trend variables. We show that the best-fitting model which minimises daily forecast errors is not the best model for forecasting peaks when the selection criterion is based on peak timing and magnitude. Furthermore, we show that best-fitting models for typical seasons contain different variables to those for atypical black swan seasons. Specifically, including seasonality in models improves performance during typical seasons but worsens it for the atypical seasons. In conclusion, we have found that including seasonality in forecast models can result in overfitting, where the models are required to be used out-of-season or during atypical seasons.