Allison Lewis, Ph.D.

Title: Applying model reduction techniques to enable efficient parameter estimation in tumor growth models

Abstract.  Numerous ODE and PDE systems have been proposed to model tumor growth dynamics, with many being subsequently used to predict future tumor behavior and inform treatment protocols.  However, the inference of model parameters for moderate- to high-dimensional systems can be challenging for several reasons: the computational cost of performing Bayesian inference in a high-dimensional space may not allow for convergence of the posterior distributions, and the full set of parameters is often unidentifiable in the sense that they are not uniquely determined by the data. This problem is exacerbated for tumor growth models, as the clinical data available for model calibration is often very sparse. The use of active subspace techniques can provide a bridge between the limited data availability and the desired model complexity by determining a subspace of identifiable inputs, then exploiting this subspace to explore only those directions in the input space that will actively inform parameter estimates. In this talk, we will demonstrate this technique on a simple—yet structurally unidentifiable—tumor growth model, and discuss how it can be utilized to inform parameter estimates for more complex ODE and PDE models.