Multi-method global sensitivity analysis of mathematical models

Increasingly-sophisticated parameter-sensitivity analysis techniques continue to be developed, and each technique comes with its own set of advantages and disadvantages. Selecting which parametersensitivity method to use for a particular model, however, is not a straightforward task. In this work, we present a multi-method framework that incorporates three global sensitivity analysis methods: two variance-based methods and one derivative-based method. The two variance-based methods are Sobol's method and MeFAST. The derivative-based method is known as DGSM (Derivative-based Global Sensitivity Measures). MeFAST (Multi test eFAST) is a new parameter sensitivity analysis implementation we built upon the eFAST (Extended Fourier Amplitude Sensitivity Test) algorithm. The improvements incorporated into MeFAST address some important aspects of prior eFAST implementations. We present an intuitive description of each implemented algorithm along with MATLAB codes and a guide to tuning algorithm hyper-parameters for better efficiency. We demonstrate the full methodology and workflow using two example mathematical models of different complexity: the first is a model of HIV disease progression and the second is a model of tumor growth. The computational framework we provide generates graphics for visualizing and comparing the results of all three sensitivity analysis algorithms (DGSM, Sobol, and MeFAST). This algorithm output comparison tool allows one to make a more informed decision when assessing which parameters most importantly influence model outcomes. Published by Elsevier Ltd.

Automated summary

This paper presents a multi-method framework that combines three global sensitivity analysis methods. The algorithms are explained with MATLAB codes and guidelines for tuning hyper-parameters. Two example models are used to demonstrate the methodology and workflow, and a graphics tool is provided for comparing the results of the sensitivity analysis algorithms.