Generative diffusion models for spatiotemporal influenza forecasting.

Influpaint, a diffusion model-based approach, accurately forecasts influenza epidemics with 30% surveillance and 70% simulated data, outperforming traditional methods in retrospective tests.

• Why it matters: Effective influenza forecasting is crucial for public health but challenging due to complex epidemic dynamics and uncertainty, highlighting the need for more flexible modeling approaches.
• What they did: The study applied denoising diffusion probabilistic models to encode influenza seasons as spatiotemporal images, using a hybrid dataset of real and simulated trajectories to learn diverse disease patterns.
• The result: Influpaint produced realistic, diverse epidemic forecasts with accuracy comparable to leading ensemble methods, demonstrating the potential of diffusion models for probabilistic infectious disease prediction.