Generative diffusion models for spatiotemporal influenza forecasting.

Area of Science:

• Epidemiology
• Artificial Intelligence
• Computational Biology

Background:

• Infectious disease forecasting is crucial for public health planning but challenging due to complex epidemic dynamics.
• Existing mechanistic and statistical models often fail to capture multimodal uncertainty and emergent trends.
• Accurate forecasting requires methods that can learn and represent complex spatiotemporal disease patterns.

Purpose of the Study:

• To introduce Influpaint, a novel approach using denoising diffusion probabilistic models for infectious disease forecasting.
• To evaluate Influpaint's ability to generate realistic and diverse epidemic trajectories.
• To assess Influpaint's forecast accuracy compared to established ensemble methods.

Main Methods:

• Influpaint encodes influenza seasons as spatiotemporal images, with pixel intensity representing disease incidence.
• It utilizes a hybrid dataset of real-world surveillance and simulated epidemic trajectories for training.
• Forecasting is performed as a conditional generation (inpainting) task using partial observational data.

Main Results:

• Influpaint successfully generates realistic and diverse epidemic trajectories.
• Retrospective evaluations show competitive forecast accuracy against leading ensemble methods.
• Real-time evaluation in the 2023-2025 U.S. CDC FluSight challenges demonstrated substantial performance improvements, though with some overconfidence in projections.

Conclusions:

• Diffusion models, like Influpaint, can effectively capture spatiotemporal structures in infectious disease dynamics.
• Influpaint offers a flexible and powerful framework for probabilistic forecasting of infectious diseases.
• Optimized training with a mix of surveillance and simulated data (30%/70%) yielded the best performance.