Back

Data-driven forecasting of Flu, RSV, and COVID-19 related outcomes in the United States and Canada via Hankel dynamic mode decomposition

Redman, W.; Mullany, L. C.

2025-11-17 infectious diseases
10.1101/2025.11.12.25339917 medRxiv
Show abstract

The (large) season-to-season variability and limited dynamical history make the forecasting of infectious diseases a challenging problem. Here, we examine the extent to which advances in data-driven dynamical modeling can provide accurate predictions by benchmarking the performance of one such method, Hankel dynamic mode decomposition (DMD), on the 2024-2025 influenza, respiratory syncytial virus (RSV), and COVID-19 seasons in the United States and Canada. Using Hankel-DMD, we generated weekly forecasts that were submitted to the Center for Disease and Controls (CDC) FluSight Forecast Hub and the University of Guelphs AI4CastingHub. Across both Hubs, we find that Hankel-DMD can provide high quality forecasts at the beginning and end of the season, but the times in-between suffer from significant overestimation of the season peak. This leads to worse than baseline performance on FluSight Forecast Hub, when submissions are evaluated across the entire season. Despite this overestimation, Hankel-DMD is found to be the best performing model for forecasting influenza, RSV, and COVID-19 in Canada, although only three other models submitted enough forecasts against which to compare. As this was the first year AI4CastingHub was active, this suggests that Hankel-DMD may be especially useful when expertise is lacking for predicting infectious dynamics in new regions. Retrospective analysis using thresholding and extensions to DMD with memory, a recently developed approach for applying DMD to non-stationary dynamical systems, provide significant improvements during the season peaks. The extent to which this can be achieved in a live forecasting setting remains to be seen. Collectively, our results demonstrate that Hankel-DMD is a promising approach for efficient and interpretable forecasting of infectious diseases and highlights a number of remaining methodological challenges which future work should aim to address.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.