Prediction and prevention of pandemics via graphical model inference and convex programming

Area of Science:

• Epidemiology
• Statistical Modeling
• Network Science

Background:

• Predicting infectious disease outbreaks like COVID-19 is challenging due to complex spatio-temporal dynamics.
• Existing statistical models often struggle to capture granular geographical spread and correlations.
• Public health requires robust methods for both predicting infection trajectories and implementing effective prevention measures.

Purpose of the Study:

• To develop algorithmic solutions for predicting the extent of infection spread in a geographical area (Inference Challenge).
• To determine the minimal intervention required to minimize the infected population (Prevention Challenge).
• To address inadequacies in current public health modeling for pandemic response.

Main Methods:

• Constructed an attractive Ising graphical model where nodes represent census tracts and edges represent interactions.
• Utilized Maximum-A-Posteriory (MAP) estimation to resolve the Inference Challenge, identifying probable infection states.
• Formulated the Prevention Challenge as a convex programming problem to find optimal control strategies.

Main Results:

• Demonstrated that attractive Ising models on dense graphs typically yield bi-modal MAP states: either full infection or no spread beyond initial nodes.
• Showcased that the Prevention Challenge can be efficiently solved using convex programming, identifying optimal prevention measures.
• Illustrated the model's efficiency on a Seattle-based simulation, revealing sparse and localized prevention strategies.

Conclusions:

• The proposed graphical model and associated algorithms offer a powerful framework for understanding and managing infectious disease outbreaks.
• Bi-modal solutions simplify prediction, enabling tractable optimization for targeted public health interventions.
• The findings suggest that optimal prevention may not always target the most highly connected areas but rather specific interaction links.