An optimized geo-hierarchical ensemble model to forecast hospitalizations from respiratory viruses in the United States

Accurate forecasting of infectious diseases is crucial for timely public health response. Ensemble frameworks have shown promising outcomes in short-term forecasting of COVID-19, among other respiratory viruses, however, there is a need to further improve these frameworks. Here, we propose the generalized Optimized Geo-Hierarchical Ensemble Model (OGEM), a novel forecasting machine learning framework to forecast state-level hospitalizations of influenza, COVID-19, and RSV in the U.S. independently. This framework is multi-resolution: it integrates state, regionally-trained, and nationally-trained models through an ensemble layer and applies various optimization methods to parameterize the model weights and enhance overall predictive accuracy. This proposed framework builds on existing forecasting literature by 1) employing an ensemble of three spatially hierarchical models with state-level forecasts as the output; 2) incorporating four distinct weight optimization methods to generate the ensemble; 3) utilizing clustering methods to dynamically identify multi-state regions as a function of short-term and long-term hospitalization trends for the regionally-trained model; and 4) providing a generalized framework to forecast the expected near-term hospitalizations from Influenza, RSV and COVID-19. Results demonstrate OGEM is a robust framework with relatively high performance. Extensive experimentation using historical data highlights the predictive power of our framework compared to existing ensemble approaches. Its robust performance underscores the framework's effectiveness and potential for improving and broadening infectious disease forecasting.