Personalization of models with many model parameters : an efficient sensitivity analysis approach

Onderzoeksoutput: Bijdrage aan tijdschriftTijdschriftartikelAcademicpeer review

27 Citaten (Scopus)
124 Downloads (Pure)

Samenvatting

Uncertainty quantification and global sensitivity analysis are indispensable for patient-specific applications of models that enhance diagnosis or aid decision-making. Variance-based sensitivity analysis methods, which apportion each fraction of the output uncertainty (variance) to the effects of individual input parameters or their interactions, are considered the gold standard. The variance portions are called the Sobol sensitivity indices and can be estimated by a Monte Carlo (MC) approach (e.g., Saltelli's method [1]) or by employing a metamodel (e.g., the (generalized) polynomial chaos expansion (gPCE) [2, 3]). All these methods require a large number of model evaluations when estimating the Sobol sensitivity indices for models with many parameters [4]. To reduce the computational cost, we introduce a two-step approach. In the first step, a subset of important parameters is identified for each output of interest using the screening method of Morris [5]. In the second step, a quantitative variance-based sensitivity analysis is performed using gPCE. Efficient sampling strategies are introduced to minimize the number of model runs required to obtain the sensitivity indices for models considering multiple outputs. The approach is tested using a model that was developed for predicting post-operative flows after creation of a vascular access for renal failure patients. We compare the sensitivity indices obtained with the novel two-step approach with those obtained from a reference analysis that applies Saltelli's MC method. The two-step approach was found to yield accurate estimates of the sensitivity indices at two orders of magnitude lower computational cost. Copyright © 2015¿John Wiley & Sons, Ltd.
Originele taal-2Engels
Pagina's (van-tot)e02727-1/18
TijdschriftInternational Journal for Numerical Methods in Biomedical Engineering
Volume3210
Nummer van het tijdschrift10
DOI's
StatusGepubliceerd - 2015

Vingerafdruk

Duik in de onderzoeksthema's van 'Personalization of models with many model parameters : an efficient sensitivity analysis approach'. Samen vormen ze een unieke vingerafdruk.

Citeer dit