Forecasting infectious disease dynamics: generative models, research attention, and the measurement of progress

Reliable forecasts of infectious disease dynamics are needed for public health action, yet transmission is complex and only partially observed. This talk explores three related questions for the future of infectious disease dynamics. First, I will present Influpaint, a generative forecasting framework that adapts denoising diffusion probabilistic models to influenza forecasting by representing epidemic trajectories as spatiotemporal images and formulating prediction as a conditional inpainting task. Trained on both surveillance and simulated data, Influpaint generates coherent epidemic scenarios and achieves competitive performance in retrospective and real-time evaluation, including the CDC FluSight challenges. Second, I will discuss how to evaluate our progress in forecasting over time. And third, I will present work on how the broader field of infectious disease dynamics shifts its attention across diseases, shaping which pathogens receive sustained research effort. Together, these projects reflect an interest in building better forecasting tools and measuring performance and progress more clearly.