Trajectory-oriented optimization of stochastic epidemiological models.

Area of Science:

• Epidemiology
• Computational Biology
• Statistics

Background:

• Epidemiological models require calibration to real-world data for accurate projections and scenario analysis.
• Stochastic models, which produce probabilistic outputs, present unique calibration challenges due to ensemble variability.
• Traditional calibration often focuses on matching mean model behavior, potentially overlooking crucial trajectory dynamics.

Purpose of the Study:

• To develop a novel calibration method for stochastic epidemiological models that accounts for random seeds.
• To ensure that calibrated model outputs, including individual trajectories, align with empirical observations.
• To improve the reliability of forward projections and what-if scenarios generated by these models.

Main Methods:

• Proposed a class of Gaussian process (GP) surrogates for efficient model exploration.
• Implemented Thompson sampling as an optimization strategy within the proposed framework.
• Introduced Trajectory Oriented Optimization (TOO) to optimize both model parameters and random seeds simultaneously.

Main Results:

• The Trajectory Oriented Optimization (TOO) approach successfully identified parameter settings and random seeds that yield model trajectories closely matching ground truth.
• This method moves beyond matching only the mean simulation behavior to capturing the dynamic realism of individual model runs.
• Demonstrated improved alignment between model outputs and empirical data compared to traditional calibration methods.

Conclusions:

• Trajectory Oriented Optimization (TOO) offers a robust method for calibrating stochastic epidemiological models.
• This approach enhances the fidelity of model simulations by ensuring individual trajectories reflect observed data.
• The findings support more accurate forecasting and scenario planning in epidemiology through improved model calibration techniques.