Content area
Accurate and timely epidemic forecasting is a cornerstone of effective public health response, especially in the early stages of pandemics like COVID-19, when reliable data are limited and situations change rapidly. This dissertation explores how artificial intelligence (AI) can make epidemic forecasting more useful, faster, and easier to scale. It focuses on two complementary directions: incorporating real-world behavioral data into statistical models, and developing deep learning–based surrogate models to speed up complex mechanistic simulations.
In Chapter 2, I examine how human mobility patterns can improve hospital admission forecasts during COVID-19. Using hospitalization data from ten hospitals in the New England region, I show that even relatively simple regression-based time series models can perform significantly better when they include mobility indicators derived from smartphone data. The findings highlight how behavioral signals can add valuable context for local, short-term decision-making during outbreaks.
Chapter 3 tackles the problem of running large-scale epidemic simulations quickly enough for real-time use. I present GLEAM-AI, a neural surrogate model trained to mimic GLEAM, a widely used stochastic agent-based simulator. Built on spatio-temporal neural processes, GLEAM-AI can produce full epidemic trajectories across geographic regions with a fraction of the original computational cost, while still providing uncertainty estimates. By conditioning on key epidemiological factors such as R0, residual immunity, and mobility trends, the model learns to replicate realistic influenza simulations without the time and resource demands of the original. This capability enables faster scenario testing, real-time forecasting, and more agile model calibration when decisions need to be made quickly.
Together, these contributions show how AI methods can enhance traditional epidemiological modeling by drawing on diverse data sources and delivering forecasts that are both scalable and probabilistic. The work provides practical tools to support public health decision-making and demonstrates the potential of deep learning to make high-fidelity epidemic modeling more accessible, adaptable, and responsive to real-world needs.