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

Area of Science:

• Epidemiology
• Computational Biology
• Machine Learning

Background:

• Accurate modeling of acute respiratory infections (ARIs) is challenging due to the multiscale nature of disease dynamics.
• Compartmental models (e.g., SIR) are simple but miss individual-level interactions.
• Individual-based models (IBMs) capture detailed transmission but are computationally intensive.

Purpose of the Study:

• To accelerate forecasting of ARI dynamics using detailed epidemic models.
• To explore hybrid and surrogate approaches for faster computation.
• To assess the accuracy and feasibility of these accelerated methods for real-time surveillance.

Main Methods:

• Developed a hybrid approach combining IBMs and compartmental models with ML-based switching.
• Implemented a surrogate approach using autoencoder approximations to replace IBM simulations.
• Compared the speed and accuracy of hybrid and surrogate methods against traditional IBMs.

Main Results:

• Hybrid approach achieved 1.6-2 times speed-up compared to standard IBMs.
• Surrogate approach yielded up to 104 times speed-up.
• Both methods maintained accuracy, making fine-grained epidemic modeling more feasible.

Conclusions:

• Machine learning and hybrid modeling significantly accelerate ARI forecasting.
• These approaches enhance the feasibility of real-time epidemic surveillance using detailed models.
• While not a full replacement, they offer practical solutions for computational challenges in epidemic modeling.