Uncertainty Quantification of Finite Element Strain Predictions for a Nitinol Medical Device: Influence of Input Parameter Probability Distribution on Output Uncertainty.

Abstract

Establishing credibility of computational modeling of medical devices is critical when an incorrect decision could cause patient harm. Uncertainty quantification (UQ) can impart credibility by providing an estimate of the model uncertainty due to variability in the input parameters. To perform UQ, non-deterministic simulations are performed where model inputs are represented by probability density functions (PDFs). Computational modeling of medical devices, however, is typically constrained by limited sample sizes and sparse experimental data. As a result, it is generally not possible to definitively characterize the true input distributions for UQ. The sparse data must instead be fit with an assumed PDF. While the assumption of a Gaussian distribution is common, other PDFs may be more appropriate depending on the context. In this study, we investigate the influence of input PDF choice on output uncertainty from non-deterministic finite element simulations of a nitinol medical device. We first characterize the geometry, material properties, and the experimental test conditions. We then perform a screening study to determine the relative importance of each parameter on our primary quantity of interest (QOI), peak strain amplitude. Next we perform three UQ studies, each using one of three different PDF types: Gaussian, gamma, and uniform. By sampling the input parameter PDFs using a Latin hypercube approach, we perform non-deterministic simulations to predict the output distributions. Our results show that use of uniform distributions yields output predictions with the largest variance and is thus the most conservative choice unless infrequent events are of interest. In contrast, we show that for conservative predictions of extreme events, a better choice is to use Gaussian or gamma distributions with asymptotic tails of finite probability that are neglected when using uniform distributions.