Leveraging synthetic and genetic data to improve epidemic forecasting

Forecasting infectious disease outbreaks is hard. Forecasting emerging infectious diseases with limited historical data is even harder. In this paper, we investigate ways to improve emerging infectious disease forecasting when little pathogen-specific training data are available. Specifically, we explore two sources of information that may be available near the start of an emerging disease outbreak: synthetic data and genetic information. For this investigation, we conducted an experiment where we trained deep learning models on different combinations of real and synthetic data, both with and without genetic information, to explore how these models compare when forecasting COVID-19 cases for US states. All models are developed with an eye towards forecasting the next pandemic. We find that models trained with synthetic data have better forecast accuracy than models trained on real data alone, and models that use genetic variants have better forecast accuracy compared to those that do not. All models outperformed a baseline persistence model, a benchmark that proved challenging for many real-time COVID-19 case forecasting models, and multiple models outperformed the COVIDHub-4_week_ensemble. This paper demonstrates the value of these underutilized sources of information and provides a blueprint for forecasting future pandemics.Author summary: Forecasting emerging infectious diseases is difficult because, by definition, little or no historical data are available for the pathogen of interest. This lack of training data is a major barrier to using flexible machine learning models in emerging disease forecasting settings. In this paper, we investigate two sources of information likely to be available near the start of a future outbreak: synthetic outbreak data and viral genetic information. Using COVID-19 as a test case, we train transformer-based models on historical respiratory disease data, synthetic outbreaks, and SARS-CoV-2 variant information to forecast weekly cases in U.S. states. Models trained with synthetic data outperform those trained on historical data alone, and models trained on both real and synthetic data perform best overall. We also find that using variant-attributable cases improves forecasts relative to forecasting total cases directly. We show these models were competitive with the best real-time COVID-19 case forecasting models. Together, these results show that synthetic data and genomic surveillance are practical, underused tools for epidemic forecasting. They offer a path toward scalable forecasting systems that can be deployed early in the next pandemic, when reliable forecasts are most needed and hardest to produce.