Bayesian inference using Gaussian process surrogates in cancer modeling

Abstract

Parametric multiscale tumor models have been used nowadays as tools to understand and predict the behavior of tumor onset, development, and decrease under treatments. In order to obtain a useful model, its parameters have to be accurately estimated, often requiring numerous model evaluations. This can be computationally prohibitive for complex problems. In this work, we propose an approximate Bayesian computation approach for estimating model parameters using a low-fidelity Gaussian Process Regression metamodel. We develop an adaptive procedure to build the data-driven surrogate model by sequentially enriching the data set in the parametric space regions where the surrogate is not accurate enough. At the end of the process, we obtain good emulators of the original models over the entire parametric space at a low computational cost. We investigate the use of the proposed framework for the calibration of two tumor growth models, reaching high accuracy and computational efficiency, which may be key issues in many complex problems.