The Susceptible-Exposed-Infectious-Removed (SEIR) model is a fundamental model in epidemiology. Model parameters such as the reciprocal transmission, incubation, and infectious rates are often difficult to measure directly, and they are estimated by solving an optimisation problem aiming to minimise the difference between the observed data and the model solution. However, the parameters of the standard SEIR system are not globally identifiable, causing optimisation algorithms to frequently converge to incorrect local optima and suffer from numerical stiffness. Here we show a comprehensive structural identifiability analysis of the SEIR framework, and present a globally identifiable and computationally stable reparameterisation of the model derived via an observational system approach. We fully characterise the multiple locally identifiable parameters, and by transforming the system into a globally identifiable structure, we eliminate the non-uniqueness issues in the parameter estimation approaches. Our numerical experiments demonstrate that this reformulation significantly improves convergence frequency, avoids runtime errors caused by numerical overflow, and consistently recovers the correct parameters. Furthermore, incorporating first-order sensitivity equations into the optimiser enhances the robustness and execution speed of the estimation process. Numerically well-conditioned methods for parameter identification, together with a comprehensive understanding of the identifiability of the parameters, ensure that the model yields reliable, rigorous insights for infectious disease forecasting and theoretical epidemiology.
🌊
