Inverse Uncertainty Quantification for Personalized Biomechanical Modeling: Application to Pulmonary Poromechanical Digital Twins.

The development of personalized models is a key step for addressing various problems, especially in biomechanics. These models typically include many constants, introduced in the model material law or loading definition, and their estimation is crucial for the model personalization. However, performing solely the estimation does not yield any information on the estimation accuracy. Additionally, all parameters can typically not be estimated based only on clinical data: some parameters are identified, while others are fixed at generic values. The question of the identifiability of the parameters, along with the robustness of the estimation, notably to noise (measurement errors) and to bias (model errors), is therefore crucial and should be quantitatively addressed in parallel to the model development. In this paper, we propose a general uncertainty quantification pipeline based on the creation of synthetic data -for which the parameters ground-truth values are known-, generated for different noise and bias levels. Estimation is then performed for many realizations of the noise or bias, as well as parameter initializations, until convergence of the estimated parameters error distributions. This pipeline was applied to a poromechanical lung model for illustration and validation purposes. It provides quantitative information on the actual identifiability of the parameters, and any derived quantity of interest, in the clinical setting. In particular, it allows to retrieve a confidence interval for each estimated parameters, which represents valuable information for diagnosis or prognosis use of the estimated values. This work is therefore a step towards improving the reliability of digital twins pipelines.