Quantifying Error Growth to Improve Infectious Disease Forecast Accuracy

PROJECT SUMMARY/ABSTRACT Over the last decade, infectious disease forecasting has advanced considerably. Using methods derived from dynamic modeling, statistical inference and numerical weather prediction, forecast systems have been developed for diseases such as influenza, SARS-CoV-2, dengue and Ebola. These systems have generated probabilistic forecasts of future epidemic outcomes with quantifiable accuracy and lead times up to 3 months, and in some instances, have been operationalized to deliver forecasts in real time. Such forecast information can be used to help manage the timing and distribution of medical countermeasures, to plan hospital and clinic staffing, and to allocate healthcare supplies in anticipation of patient surges. Ongoing research is needed to further improve the accuracy of these disease forecasts so that the decisions and actions that are based on this information are more soundly motivated. To this end, it is vital that the sources of error in infectious disease forecasts are better understood, that the growth of error during forecast is quantified, and that methods are developed to control and optimize that error growth in order to improve forecast accuracy. The aim of this project is to leverage methods that have been employed to understand and quantify error growth in weather forecasting models and to improve weather forecasting accuracy, and to apply these methods to infectious disease forecasting systems. Specifically, we will: 1) quantify the nonlinear growth of error within a diversity of infectious disease forecasting models and then develop methods to optimize that error growth during forecasting, thus improving forecast accuracy; we hypothesize that the fastest growing mode within disease forecasting models can be identified using singular vector analysis (SVA); quantified error growth can then be exploited using optimal perturbation methods, in conjunction with observations and data assimilation approaches, to generate a more calibrated ensemble forecast that produces more accurate probabilistic predictions; 2) apply SVA and optimal perturbation methods to a recently validated, spatially explicit model of influenza in order to understand how uncertainty propagates when observations are missing and to identify which locations are critical for accurate forecasting throughout the network; we hypothesize these findings can be used to identify improved, more optimal disease surveillance networks; and 3) develop models to forecast and project the continued spread of influenza and SARS-CoV-2 internationally; here, we will develop multi- country spatially-explicit networked metapopulation models capable of accurate simulation and forecasting of the transmission and spread of seasonal influenza and SARS-CoV-2 within and between countries; we hypothesize that the intra- and inter-country spread of these diseases can be forecast more accurately with systems that utilize network model structures. The findings from this project will improve understanding of error growth in forecast models, improve the accuracy of operational infectious disease forecasting, inform surveillance practices, and enable more accurate forecast of the spread of disease.