Uncertainty Quantification in Mechanistic Epidemic Models
Résumé
This work introduces a novel data-driven approach for parameter estimation and uncertainty quantification of epidemic models. The proposed methodology combines two key elements: (i) the identification of initial conditions through plausible dynamic states that align with observed data, enabling a more accurate representation of the epidemic's initial state; (ii) the integration of a cross-entropy-based method to learn an informative prior distribution for the model parameters, enhancing the reliability of parameter estimation. The employed epidemic model features a generalized SEIR structure with time-dependent transmission rates, asymptomatic cases, and hospitalizations. A multi-objective optimization problem is formulated, considering two cost terms related to hospitalizations and deaths, leading to the identification of twelve model parameters. The calibrated model not only provides a consistent description of the available data but also enables short-term forecasts, offering valuable real-time insights for epidemic modeling. This data-driven approximate Bayesian computation framework contributes to improved parameter estimation, uncertainty quantification, and forecasting capabilities, facilitating evidence-based decision-making and enhancing the overall understanding and management of epidemic outbreaks.