Bayesian sensitivity analysis of a 1D vascular model with Gaussian process emulators

One‐dimensional models of the cardiovascular system can capture the physics of pulse waves but involve many parameters. Since these may vary among individuals, patient‐specific models are difficult to construct. Sensitivity analysis can be used to rank model parameters by their effect on outputs and to quantify how uncertainty in parameters influences output uncertainty. This type of analysis is often conducted with a Monte Carlo method, where large numbers of model runs are used to assess input‐output relations. The aim of this study was to demonstrate the computational efficiency of variance‐based sensitivity analysis of 1D vascular models using Gaussian process emulators, compared to a standard Monte Carlo approach. The methodology was tested on four vascular networks of increasing complexity to analyse its scalability. The computational time needed to perform the sensitivity analysis with an emulator was reduced by the 99.96% compared to a Monte Carlo approach. Despite the reduced computational time, sensitivity indices obtained using the two approaches were comparable. The scalability study showed that the number of mechanistic simulations needed to train a Gaussian process for sensitivity analysis was of the order O(d), rather than O(d×103) needed for Monte Carlo analysis (where d is the number of parameters in the model). The efficiency of this approach, combined with capacity to estimate the impact of uncertain parameters on model outputs, will enable development of patient‐specific models of the vascular system, and has the potential to produce results with clinical relevance. A statistical emulator based on Gaussian process regression is used to predict 1D vascular model outcomes needed to perform the sensitivity analysis. Four vascular models of increasing size and complexity are used to test this methodology. The number of training samples needed to train the Gaussian process emulator is proved to be 3 order of magnitudes smaller than the number of 1D simulation runs needed for a traditional Monte Carlo analysis. The scored mean average prediction error is maintained below 1%.