Epidemiological model calibration via graybox Bayesian optimization.

Area of Science:

• Epidemiology
• Computational Biology
• Statistical Modeling

Background:

• Traditional epidemiological model calibration methods assume low computational cost, which is often not feasible for complex models.
• There is a need for efficient calibration techniques that can handle computationally expensive epidemiological models.

Purpose of the Study:

• To develop efficient calibration methods for compartmental epidemiological models using Bayesian decision-making.
• To introduce a "graybox" Bayesian optimization (BO) scheme that leverages the functional structure of epidemiological models for enhanced calibration.
• To propose decoupled decision-making strategies within BO to further exploit model structure.

Main Methods:

• Utilizing Gaussian processes as surrogates for computationally expensive epidemiological models.
• Implementing a "graybox" Bayesian optimization framework tailored for compartmental models.
• Developing decoupled decision-making strategies for BO to enhance calibration efficiency.

Main Results:

• The proposed graybox BO schemes efficiently calibrate computationally expensive epidemiological models.
• Improved calibration performance was observed, measured by the logarithm of mean squared errors.
• Faster convergence of performance was achieved in terms of BO iterations.

Conclusions:

• The developed graybox Bayesian optimization methods offer efficient calibration for complex epidemiological models.
• These methods enhance calibration performance and speed, particularly for computationally intensive models.
• The approach shows potential for extension to even more complex models like agent-based models.