Integrating Machine Learning with Hybrid and Surrogate Models to Accelerate Multiscale Modeling of Acute Respiratory Infections

Accurate, efficient, and explainable modeling of the dynamics of acute respiratory infections (ARIs) remains, in many aspects, a significant challenge. While compartmental models such as SIR (Susceptible-Infected-Recovered) remain widely used for that purpose due to their simplicity, they cannot capture the complicated multiscale nature of disease progression which unites individual-level interactions affecting the initial phase of an outbreak and mass action laws governing the disease transmission in its general phase. Individual-based models (IBMs) offer a detailed representation capable of capturing these transmission nuances but have high computational demands. In this work, we explore hybrid and surrogate approaches to accelerate forecasting of acute respiratory infection dynamics performed via detailed epidemic models. The hybrid approach combines IBMs and compartmental models, dynamically switching between them with the help of statistical and ML-based methods. The surrogate approach, on the other hand, replaces IBM simulations with trained autoencoder approximations. Our results demonstrate that the usage of machine learning techniques and hybrid modeling allows us to obtain a significant speed-up compared to the original individual-based model-up to 1.6-2 times for the hybrid approach and up to 104 times in case of a surrogate model-without compromising accuracy. Although the suggested approaches cannot fully replace the original model, under certain scenarios they make forecasting with fine-grained epidemic models much more feasible for real-time use in epidemic surveillance.