An Ensemble Method for Disease Surveillance

Healthcare heavily relies on effective disease surveillance, particularly during epidemiological emergencies. Predicting trends is challenging but crucial for effective policies. We propose an ensemble modelling framework to forecast time series of counts, such as the number of cases or hospitalizations. The method relies on multiple models, including flexible regressions with cubic splines, generalized additive models, Bayesian structural time series, and integer-valued generalized autoregressive conditional heteroscedasticity models. Predictions are computed as weighted averages on some of the individual models’ predictions, where model selection and weights are based on performance on past observations. This approach allows to exploit diverse model strengths and compensate for individual weaknesses, ensuring more stable predictions. The behavior of the proposed method is illustrated on daily COVID-19 hospitalization counts in Emilia-Romagna in 2022, with a focus on periods showing changing trends.